ROADWAYS, INPUT SOURCING, AND PATTERNS OF SPECIALIZATION
Esteban Jaimovich
LATIN AMERICAN AND THE CARIBBEAN ECONOMIC ASSOCIATION
June 2018
The views expressed herein are those of the authors and do not necessarily reflect the views of the
Latin American and the Caribbean Economic Association. Research published in this series may
include views on policy, but LACEA takes no institutional policy positions.
LACEA working papers are circulated for discussion and comment purposes. Citation of such a
paper should account for its provisional character. A revised version may be available directly from
the author.
© 2018 by Esteban Jaimovich. All rights reserved. Short sections of text, not to exceed two
paragraphs, may be quoted without explicit permission provided that full credit, including © notice,
is given to the source.
LACEA WORKING PAPER SERES. No. 0008
LACEA WORKING PAPER SERIES No. 0008 June 2018
Roadways, Input Sourcing, and Patterns of Specialization
Esteban Jaimovich
University of Surrey
ABSTRACT
We propose a model where the internal transport network facilitates the sourcing of intermediate
goods from different locations. A denser internal transport network promotes thus the growth of
industries that rely on a large variety of inputs. The model shows that heterogeneities in internal
transport infrastructures can become a key factor in shaping comparative advantage and
specialization. Moreover, when sufficiently pronounced, such heterogeneities may even
overshadow more traditional sources of specialization based on factor productivities. Evidence
based on industry-level trade data grants support to the main prediction of the model: countries with
denser road networks export relatively more in industries that exhibit broader input bases. We show
that this correlation is robust to several possible confounding effects proposed by the literature, such
as the impact of institutions on specialization in complex goods. Furthermore, we show that a
similar correlation arises as well when the density of the local transport network is measured by the
density of their internal waterways, and also when road density is instrumented with measures of
terrain roughness.
JEL Classification: F11, O18, R12
Keywords: International Trade, Comparative Advantage, Internal Transportation Costs.
1 Introduction
The spatial distribution of economic activities means that transportation costs represent a
major factor in�uencing countries�output, trade �ows and specialization. Apart from few ex-
ceptions, the vast majority of the past trade literature has centered their attention on the cost
of shipping goods internationally.1 However, the evidence at hand suggests that internal trans-
port costs are far from being a secondary component that can be disregarded when confronted
with transboundary costs.2 Furthermore, the impact of internal transport costs on special-
ization gets magni�ed by the fact that local infrastructures di¤er quite substantially between
countries, especially when comparing economies at di¤erent stages of development.
Being able to e¢ ciently transport commodities across space is crucial to keep total costs low.
Yet, owing to speci�cities of their physical characteristics and of their production processes,
some commodities turn out to be inherently more transport-intensive than others. This means
that the e¢ ciency of the local transportation infrastructure may unevenly a¤ect the devel-
opment of di¤erent industries. This paper studies a speci�c channel by which the internal
transport network may shape countries�comparative advantages and specialization. One key
role of the internal transportation network is that it facilitates the sourcing of intermediate in-
puts from di¤erent locations. As a result, industries that require a large variety of intermediate
inputs tend to make more intense use of the network.3
To illustrate this idea, we introduce a simple model with two intermediate inputs and a
continuum of �nal good producers. A denser road network allows cheaper transportation of the
intermediate inputs to the location site of �nal good producers. A crucial feature of the model
is that industries producing �nal goods di¤er in terms of the breadth of their intermediate
input requirements. In particular, some industries have production functions that are very
intensive in only one intermediate input, while others require a more balanced mix of the two
intermediate inputs. Since transportation of inputs is costly, those industries that require a
relatively balanced combination of the intermediate inputs turn out to bene�t relatively more
(in terms of cost reduction) from a denser road network.
1For a few papers that have incorporated internal transport costs into trade models, see Allen and Arkolakis
(2014), Cosar and Fajgelbaum (2016), Ramondo, Rodriguez-Clare and Saborio-Rodriguez (2016), Redding
(2016), Matsuyama (2017).2See, e.g., Limao and Venables (2001), Anderson and Van Wincoop (2004), Hillberry and Hummels (2008),
Mesquita Moreira et al (2013), Agnosteva et al (2014), Atkin and Donaldson (2015), Donaldson (2018).3This idea was �rst suggested by Clague (1991a, 1991b) who argued that countries with poor infrastructure
will specialize in �self-contained�sectors (i.e., sectors that do not intensively rely on inputs from other sectors).
2
This simple mechanism yields a very clear prediction in terms of specialization within a
framework with open economies. Countries that enjoy a denser local transport network tend
to display a comparative advantage in the goods whose production process requires a relatively
balanced mix of the intermediate inputs. This is because these are the industries that make
heavier use of the local transport network to source their inputs. Conversely, countries with
underdeveloped transport networks tend to specialize in industries with narrow input bases, as
this allows them to economize on input sourcing.
After presenting the model we provide evidence consistent with its main prediction. To do
so, we proceed as follows. Firstly, we index industries by their degree of input breadth using
the information contained in the US input-output matrix. Secondly, we measure the density
of local transport networks of countries by the length of their roadways per square kilometer.
Finally, we correlate countries specialization by industries (measured by their total exports at
the industry level) with an interaction term between industries�input breadth and countries�
roadways density. We �nd that countries with denser road networks export relatively more in
industries that exhibit broader input bases.
The correlation between road density and specialization in industries with broader input
bases may obviously be driven by other mechanisms to the one suggested by our model. We
show however that this correlation is robust to the inclusion of a large set of possible confounding
covariates. In particular, one important channel related to ours works through institutions, as
industries that rely on a wide set of inputs tend to be more dependent on contract enforcement
[Levchenko (2007) and Nunn (2007)]. We show that the correlation predicted by our model
is still present once we also control for the e¤ect of institutions. In that respect, our �ndings
complement the previous studies that have interpreted the degree of input variety as a sign of
product complexity, showing that industries with wide input bases seem also to be strongly
reliant on the internal transport network.
One additional concern is whether the found correlation can be interpreted at all as evidence
of causation from road density to specialization in transport-intensive industries. Roadways are
the result of investment choices. Hence, road infrastructure may positively respond to transport
needs resulting from patterns of specialization, reversing thus the direction of causation. Inter-
estingly, we show that an analogous correlation to that one found with road density arises when
using waterways density as an alternative measure of the depth of the local transport network.
Moreover, this correlation is especially strong and signi�cant in the case lower-income countries,
which are exactly the types of economies that tend to su¤er from sparser road networks.
Arguably, while waterways cannot be molded and expanded as �exibly as road networks,
3
and hence they are less sensitive to issues of reverse causation, their evidence does not directly
address this concern. In order to address more directly the possibility that reverse causality is
behind our empirical results, drawing on Ramcharan (2009), we also instrument the density of
a country�s road network with topographical measures of terrain roughness.4 The instrumental
variable approach con�rms the previous �ndings, granting further support to the hypothesis
that the density of the internal road network is an important determinant of comparative
advantage in industries with wide input bases.
There is a growing literature studying the impact of the local transport infrastructure on
international and intra-regional trade and specialization. For example, Volpe Martincus and
Blyde (2013) study the access to foreign markets and international trade across regions in
Chile, Cosar and Demir (2016) does so for Turkey, and Volpe Martincus, Carballo and Cusolito
(2017) for Peru. Donaldson (2018) looked at reductions of price and output distortions across
Indian regions after expansions of the local railroad network, and Donaldson and Hornbeck
(2016) assess how the expansion of the railroad network in the US enhanced market access of
US counties. Fajgelbaum and Redding (2014) and Cosar and Fajgelbaum (2016) investigate
the regional location of export-oriented activities given the local infrastructure in the cases of
Argentina and China, respectively. Closer to our main focus, Duranton, Morrow and Turner
(2014) and Cosar and Demir (2016) have tried to capture whether there is some e¤ect of road
infrastructure on specialization in transport-intensive activities. Duranton et al (2014) show
that US cities with more highways tend to produce goods of higher weight per physical unit,
while Cosar and Demir (2016) �nd a similar e¤ect for Turkey. Our paper focuses on a di¤erent
channel whereby the local transport infrastructure impacts comparative advantages: the notion
that the spatial distribution of activities makes industries that need to source a large variety
of intermediate inputs relatively more reliant on the internal transport network.
The internal transportation channel studied in this paper was �rst suggested by Clague
(1991a, 1991b). There it is argued that poorer economies specialize in �self-contained�sectors,
as they lack a su¢ ciently developed infrastructure needed to sustain the production of industries
that require a large variety of inputs. These articles, however, do not articulate this hypothesis
within an international trade model, nor do they empirically assess whether trade �ows at
the industry level are associated with actual measures of the internal transport network in a
way consistent with it.5 We formulate the hypothesis that the local transport infrastructure
4Ramcharan (2009) shows that countries with rougher topography tend to exhibit less dense road networks.
He argues that this is partly due to the impact of terrain roughness and grade variation on the cost of building
and maintenance of transport networks.5These articles provide evidence that the relative e¢ ciency of underdeveloped economies is worse in industries
4
matters relatively more for industries with wider input bases within a trade model, where
transport costs, location choices and comparative advantage are explicitly modeled. This leads
to an endogenous determination of trade �ows and specialization patterns, which respond to
heterogeneities in transport infrastructures. In addition, we present evidence supporting the
relevance of this mechanism exploiting cross-country variation in the density of road networks.6
Finally, our paper also relates to several strands of literature that have expanded upon the
traditional Ricardian/Heckscher-Ohlin trade models based on heterogeneities in factor produc-
tivities/endowments. One set of papers have looked at enforcement institutions as a source of
comparative advantage in industries producing complex goods requiring large variety of input-
speci�c relationships [Antràs (2005), Acemoglu, Antràs and Helpman (2007), Levchenko (2007),
Nunn (2007), Costinot (2009), Ferguson and Formai (2013)]. Another strand of literature has
delved into the role of �nancial markets fostering exports in industries that are heavy users
of external �nance [Beck (2002), Svaleryd and Vlachos (2005), Becker, Chen and Greenberg
(2012), Manova (2013)]. Finally, another institutional source of comparative advantage is pre-
sented by Cuñat and Melitz (2012), who show that countries with more �exible labor market
regulations tend to export more in industries subject to higher volatility.7 Our paper high-
lights the impact of local infrastructures when industries di¤er in their dependence on internal
transportation of inputs.
The rest of the paper is organised as follows. Section 2 introduces the main features of
the model in the case of a closed economy. Section 3 extends the model to a two-country
setup, and derives the main predictions in terms of comparative advantage and trade �ows.
Section 4 contrasts the main predictions of the model with the data. Section 5 discusses
some endogeneity issues and alternative interpretations of the empirical results. Section 6
discusses the empirical plausibility some of the key assumptions implicit in the model in terms
of geographic distribution of industries. Section 7 concludes.
that rely on a large variety of intermediate inputs. While this could be the result of poorer economies having less
developed transport networks, it could also be the result of other factors usually associated with underdeveloped
economies, like weaker institutions, lower levels of human capital, etc.6Yeaple and Golub (2007) show that the stock of roads a¤ects total factor productivity and sectoral com-
position across 10 industries for a panel of 18 countries. While their analysis highlights that roads may be a
source of comparative advantage in some industries, it does not link the e¤ect of heterogeneities in transport
infrastructures to specialization in industries with di¤erent degrees of input diversity.7See Chor (2010) for a paper that aims at quantifying the importance of all these institutional sources of com-
parative advantage, alongside the more traditional ones stemming from heterogeneities in factor productivities
and endowments.
5
2 General Setup in a Closed Economy Model
This section presents the environment and main features of our model in the speci�c case of
a closed economy. Starting o¤ with a closed economy proves helpful in two aspects. First,
it allows an easier description of the main building blocks of the model. Second, it facili-
tates the exposition of the main intuition for how the density of the transport network may
heterogeneously a¤ect the cost of production in di¤erent sectors.
2.1 Intermediate and Final Goods Sector
There exists a unit continuum of �nal goods, indexed by j 2 [0; 1]. All �nal good markets areperfectly competitive. Final goods are purchased by individuals with preferences given by
U =
Z 1
0
ln(yj) dj; (1)
where yj denotes the consumed amount of j. There is a mass of individuals equal to L. Each
individual is endowed with one unit of labor which is supplied inelastically for a wage w.
In addition to the set of �nal goods, there exist two intermediate goods, indexed by i = 0; 1.
There is free entry to the markets of both intermediate goods. Each intermediate good is
produced with labor, according to the following linear production functions:
Xi =Li1 + "i
, i = 0; 1: (2)
In (2), Xi denotes the total amount of intermediate good i produced in the economy, Li is the
total amount of labor used in producing i, and "i � 0 is a technological parameter determininglabor productivity in sector i.8
Final goods are produced by combining the two intermediate goods within Cobb-Douglas
production functions. Total output of �nal good j 2 [0; 1] is given by:
Yj =1
��jj (1� �j)1��j
X1��j0;j X
�j1;j, where �j 2 [0; 1], (3)
and X0;j and X1;j denote the amount of intermediate good 0 and 1 used in the production of
�nal good j, respectively.
The Cobb-Douglas production functions (3) di¤er across �nal good sectors in terms of the
intensity requirements of each intermediate good. Sectors with a small (resp. large) �j use
8The model can be generalized to comprise N di¤erent intermediate goods. Appendix D brie�y presents an
example with three intermediate goods.
6
input 0 (resp. input 1) more intensively. On the other hand, sectors whose �j lies in the
vicinity of 0:5 tend to use a relatively balanced mix of both inputs. For the rest of the paper,
we will assume that, when considering the whole set of �nal good producers, the values of
�j are uniformly distributed within the unit interval. Abusing a bit the notation, we can thus
henceforth index �nal goods by their value of �j.9
Perfect competition in �nal good markets implies that, in equilibrium, each �nal good j will
be sold at a price equal to its marginal cost. Using (3), we can obtain the expression for the
marginal cost, which we denote by cj. Namely,
cj = p1��j0;j p
�j1;j; (4)
where p0;j and p1;j are the prices at which the producer of �nal good j can purchase each unit
of input 0 and 1, respectively.10
2.2 Geographic Structure of the Economy
We assume that each intermediate good is produced in a di¤erent site, which we refer to as
location 0 (for input 0) and location 1 (for input 1). Labor is perfectly mobile across locations
at zero cost. Intermediate goods must, however, incur an iceberg transport cost to be moved
around. When the distance between the location of j and that of i is dj;i � 0, the intermediategood producer i must ship 1 + t dj;i units of input i in order for the �nal good producer j to
receive one unit of i.11
There exists a road network of length r linking location 0 and location 1. We assume that
the shortest distance between location 0 and 1 is given by a function '(r), with '0(r) < 0.
That is, we assume that longer road networks facilitate transportation across location 0 and9None of the main results in the model strictly depend on the Cobb-Douglas speci�cation, and we could
alternatively use a more general CES production function, where: Yj = [(1� �j) X�0;j + �j X
�1;j ]
1=�, with
�j 2 [0; 1] and � < 1: Notice that this excludes the trivial case of a linear production function in X0;j and X1;j(i.e., � = 1), as this would imply perfect substitutability between the inputs. Appendix D shows how the main
results in this section remain true with a general CES function. In the end, the choice of (3) is essentially owing
to its algebraic neatness. In addition, in the Cobb-Douglas case, the weights �j and 1 � �j carry a very clearinterpretation: they are always equal to the share of each intermediate input over the total cost of intermediates.10Although we are assuming that there is free entry in the intermediate goods sectors, in principle, our model
will not always lead to the same price paid by each �nal good producer j for each of the inputs. The reason
for this is that both p0;j and p1;j will also incorporate internal transport costs, and these costs may well di¤er
across �nal good producers given their location and the locations of intermediate goods.11Appendix D shows that all the main results of this section would remain essentially intact if we assumed
that the transport cost on inputs is additive instead of multiplicative.
7
1 by shortening the distance between the two locations. In Appendix A, we provide a simple
geographical structure of the economy as illustration of the function '(r) and the fact that is
strictly decreasing in r.
2.3 Location Choice by Final Good Producers
The previous subsection assumed that each intermediate good is produced in a speci�c and
exogenously given location. With regards to �nal goods producers, we assume that they can
freely choose a location on any point along the road network linking location 0 and location 1.
Given that shipping inputs across production sites entails a transport cost, �nal good pro-
ducers will choose their own location so as to minimize their marginal costs (cj). Recall that,
given a road network of length r, the distance between location 0 and 1 is equal to '(r). Let
now lj'(r) denote the (minimum) distance between the location chosen by producer j and
location 0, where lj 2 [0; 1]. Notice lj = 0 means that j selects location 0, while lj = 1 meansthat j chooses location 1. On the other hand, interior values of lj �that is, lj 2 (0; 1)�entailthat j locates itself at somewhere along the road network linking location 0 and 1.
Given the selected lj 2 [0; 1], producer j must thus pay
p0;j = [1 + lj'(r)t] (1 + "0)w
for each unit of input 0 that he purchases, while he must pay
p1;j = [1 + (1� lj)'(r)t] (1 + "1)w
for each purchased unit of input 1.
Bearing in mind (4), producer j will thus choose his location by solving:
minlj2[0;1]
: cj(lj) = [(1 + lj'(r)t) (1 + "0)w]1��j [(1 + (1� lj)'(r)t) (1 + "1)w]�j : (5)
The above problem yields corner solutions. Comparing thus cj(0) vis-a-vis cj(1), we obtain
l�j =
8<: 0 if �j � 0:5
1 if �j � 0:5(6)
The expression in (6) represents an intuitive agglomeration result: �nal producers choose to
locate their �rm in the same place where the input they use more intensively is being produced.12
12Corner solutions in (5) stem from the fact that transport costs are assumed linear in distance. If transport
costs were convex in distance the model could yield interior solutions (at least for those j with values of �j that
lie near one half). In any case, even with su¢ ciently convex distance costs, the same qualitative patterns of
agglomeration between intermediate and �nal producers would obtain: sectors with relatively small �j (resp.
large �j) will locate relatively closer to location 0 (resp. location 1).
8
Finally, plugging (6) back into the expression in the right-hand side of (5) we can obtain
the marginal cost of �nal good j:
c�j =
8<: (1 + '(r)t)�j (1 + "0)1��j (1 + "1)
�j w if �j � 0:5
(1 + '(r)t)1��j (1 + "0)1��j (1 + "1)
�j w if �j � 0:5(7)
The expressions in (7) shows that the marginal cost of �nal good j is determined by the
labor cost of producing the required inputs (via the wage w, and the parameters "0 and "1), and
also by the transport cost involved in sourcing those inputs. Importantly, recall that �nal good
producers will optimally choose to set up their �rms in the same location where the input they
use more intensively is being produced. As a result, the transport cost ends up being applied
only to the input whose Cobb-Douglas weight in (3) is smaller than 0:5. In turn, this implies
that internal transport costs tend to a¤ect more severely the marginal cost of those �nal goods
whose �j lies near 0:5. In other words, internal transport costs tend to particularly hurt sectors
which use a relatively balanced combination of inputs. On the other hand, this also implies
that while improvements in transport infrastructure will lower the cost of production of all �nal
goods (except for the extreme cases where either �j = 0 or �j = 1), such improvements will
end up lowering the marginal cost of goods whose �j is closer to 0:5 by relatively more. The
following lemma states this result more formally.
Lemma 1 Consider the expression for the marginal cost of good j in (7) and two generic values
of the road length r1 and r2, such that r1 < r2. Then,
1. c�j(r1)=c�j(r2) > 1 for all �j 2 (0; 1), while c�j(r1)=c�j(r2) = 1 when �j = 0 and �j = 1:
2. The ratio c�j(r1)=c�j(r2) is strictly increasing in �j for all �j 2
�0; 1
2
�and strictly decreasing
in �j for all �j 2�12; 1�. Moreover, the highest value of c�j(r1)=c
�j(r2) is reached at �j =
12.
Lemma 1 shows that larger values of r lead to lower marginal costs of production, but that
the fall in the marginal cost is proportionally greater in sectors with values of �j closer to 12.
In the next section, where we extend the model to allow international trade and specialization,
this result will turn the density of the road network into a source of comparative advantage:
countries with denser road networks will tend to enjoy a comparative advantage in sectors with
intermediate levels of �j.
9
3 Two-Country Model
We consider now a world economy a lá Dornbusch-Fischer-Samuelson (1977) with two countries:
H and F . Both countries are populated by a mass L of individuals. Each individual is endowed
with one unit of labor that is supplied inelastically in the local labor market. We let wH and
wF denote the wage in H and F , respectively. Henceforth, we set wF = 1 (i.e., we set wF as
the numeraire), and use ! � wH=wF to denote the relative wage. All individuals share the
same preferences �given by (1)�over the unit continuum of �nal goods.
Each �nal good could in principle be produced by any of the two countries. The technologies
to produce �nal goods are identical in both H and F , given by the Cobb-Douglas functions (3).
All �nal goods markets are perfectly competitive. In addition, we assume that all �nal goods
are internationally tradeable, subject to an iceberg cost � > 0 (that is, when 1 + � units of j
are shipped internationally, only 1 unit of j will arrive at the destination country).
Unlike for �nal goods, we assume that intermediate goods are non-tradeable internation-
ally.13 We also assume that the technologies to produce the intermediate goods di¤er between
H and F . Letting Xi;c denote the total amount of intermediate good i produced in country c,
we assume that in H
X0;H = L0;H and X1;H =L1;H1 + "
; (8)
while in F ,
X0;F =L0;F1 + "
and X1;F = L1;F ; (9)
where Li;c is the total amount of labor used in producing input i in country c, and " > 0. There
is free entry to the intermediate goods markets in both H and in F .
Two features implied by (8) and (9), coupled with the �nal goods production functions (3),
are worth stressing. First, since they imply that H is relatively more productive than F in
sector i = 0, they tend to yield a comparative advantage by H on the �nal goods that rely more
heavily on input 0 (that is, on those j whose �j is small). Second, since (8) and (9) exactly
mirror one another, they implicitly assume away any aggregate absolute advantage by one
country over the other one stemming from the distribution of sectoral labor productivities.14
13Restricting international trade only to �nal goods simpli�es the exposition of the main results of the model.
In principle, we could allow for trade of intermediates as well, as long as (analogously to the case of domestically
produced inputs) imported inputs need, to some extent, to be transported internally until reaching the exact
location of domestic �nal good producers.14The model could be generalized to encompass production functions Xi;c = Li;c=(1 + "i;c), where, i = 1; 2,
c = H;F and "i;c � 0. We deliberately choose a symmetric distribution of labor productivities, as is (8) and
(9), since it allows a cleaner depiction of the impact of road networks on the patterns of comparative advantage.
10
Analogously to the closed economy setup in Section 2.2, we assume that each input is
produced in a speci�c location. We keep referring as location 0 to the production site of input
0, and as location 1 to that one of input 1. (In this case, there is one such location in each
of the countries.) Also like in the closed economy setup, we assume that the distance between
location 0 and 1 in country c depends on the length of the road network in c via the distance
function '(rc). We also assume that the iceberg cost t per unit of distance dj;i travelled by
input i to reach producer j is identical in H and F .15
We denote now by rH and rF the length of the road network in H and F , respectively.
Henceforth, we assume:
Assumption 1 rF < rH :
In our model, Assumption 1 will convey a source of comparative advantage toH in the types
of goods that depend on (internal) transport of inputs more strongly. In addition, rH > rF also
implies that H can, in general, ship inputs internally at lower cost than F . This fact will in
turn grant a source of aggregate absolute advantage by H over F .
3.1 Pricing of Final Goods in H and F
The fact that all good markets in H and F are perfectly competitive implies again that �nal
goods will be sold at their marginal costs. Notice that this will include both the incurred
internal and international transport costs. In its general form, the price of �nal good j 2 [0; 1]produced in country c = H;F and sold in country m = H;F will be given by
Pmj;c = (1 + � � Im6=c) [(1 + lj;c '(rc)t) (1 + "0;c)]1��j [(1 + (1� lj;c)'(rc)t) (1 + "1;c)]�j wc; (10)
where: i) Im6=c is an indicator function that is equal to one when m 6= c, and zero otherwise; ii)"0;H = "1;F = 0 and "1;H = "0;F = "; iii) lj;c '(rc), where lj;c 2 [0; 1], is the (minimum) distancebetween producer j in country c and location 0.
Final good producers will optimally seek to minimize their marginal costs. Analogously as
done in Section 2.3, it can be proved that this is achieved by setting up �rm j in location 0
15One may �nd it somewhat arti�cial the fact that the model assumes that transporting intermediate goods
within the country entails a cost, but at the same time it abstracts from any cost regarding transportation of
�nal goods. Yet, in the context of our model, none of the main results would be qualitatively altered by adding
an internal transport cost of �nal goods, provided this cost applies both to locally produced and imported goods,
and that the direct cost of transportation for �nal goods bears no systematic relation with the parameter �j .
11
when �j � 0:5, and setting it up in location 1 when �j � 0:5. That is, condition (6) still holdstrue within the two-country model, with lj;c = l�j for c = H;F .
By using this result, together with (10), the price of good j when produced in country H
and sold in m = H;F , denoted by Pmj;H , can be written as
Pmj;H =
((1 + � � Im6=H) (1 + '(rH)t)�j (1 + ")�j ! if �j � 0:5(1 + � � Im6=H) (1 + '(rH)t)1��j (1 + ")�j ! if �j � 0:5
: (11)
Analogously, Pmj;F , with m = H;F , can be written as
Pmj;F =
((1 + � � Im6=F ) (1 + '(rF )t)�j (1 + ")1��j if �j � 0:5(1 + � � Im6=F ) (1 + '(rF )t)1��j (1 + ")1��j if �j � 0:5
: (12)
To ease notation, it proves convenient to de�ne
� � 1 + '(rF )t
1 + '(rH)t: (13)
Notice that � > 1, since rF < rH . In the context of our model, � can be interpreted as a
measure of the advantage of H over F in terms of length of road network.
3.2 Traded (and Non-Traded) Goods
In equilibrium, consumers will buy each �nal good j from the producer who can o¤er j at the
lowest price. In some cases this will mean that consumers will source good j locally, while in
others they will choose to import it. Naturally, given that shipping �nal goods internationally
entails an iceberg cost � > 0, if in equilibrium country c is an exporter of good j, then it must
also be the case that individuals from c must be buying good j from local producers.
By comparing (11) vis-a-vis (12), we can observe that international trade of �nal goods takes
place when the following conditions hold true (henceforth, without any loss of generality, we
assume that when confronted with identical prices, consumers always buy from local producers).
� H will export �nal good j to F if and only if:
! < (1 + �)�1 ��j (1 + ")1�2�j when �j � 0:5;
! < (1 + �)�1 �1��j (1 + ")1�2�j when �j � 0:5(14)
� H will import �nal good j from F if and only if:
! > (1 + �) ��j (1 + ")1�2�j when �j � 0:5;
! > (1 + �) �1��j (1 + ")1�2�j when �j � 0:5(15)
12
The presence of � > 0 in (14) and (15) implies that some �nal goods may end up not being
traded internationally. In particular, if for some subset of �nal goods whose 0 � �j � 0:5,
the model yields (1 + �)�1 � !���j (1 + ")2�j�1 � (1 + �), then consumers from both H and
F will end up sourcing these goods locally. Similarly, if for some subset of �nal goods whose
0:5 � �j � 1, the model yields (1 + �)�1 � !��j�1 (1 + ")2�j�1 � (1 + �), these goods will alsobe sourced in both H and F from local producers.
3.3 Equilibrium and Patterns of Specialization
In equilibrium, the total (world) spending on �nal goods produced in each country must equal
the total labor income of each country. In our two-country setup, this condition can be restated
as a trade balance equilibrium for eitherH or F . The utility function (1) implies that consumers
allocate identical expenditure shares across all �nal goods in the optimum.16 Hence, in our
model, the equilibrium condition in the world economy boils down to:Z 12
0
I�! < (1 + �)�1 ��j (1 + ")1�2�j
d�j +
Z 1
12
I�! < (1 + �)�1 �1��j (1 + ")1�2�j
d�j =
"Z 12
0
I�! > (1 + �)��j (1 + ")1�2�j
d�j +
Z 1
12
I�! > (1 + �)�1��j (1 + ")1�2�j
d�j
#!;
(16)
where I f�g in (16) is an indicator function that is equal to 1 when the condition inside thebraces holds true, and 0 otherwise. The left-hand side of (16) thus amounts to the total value
of H�s exports, whereas its right-hand side equals the total value of H�s imports.
Henceforth, we impose an additional parametric restriction to the model:
Assumption 2 " > �:
Assumption 2 ensures that our model will always feature positive trade in equilibrium. Intu-
itively, " > � implies that the source of comparative advantages linked to heterogeneities in
sectoral labor productivities �i.e., those determined by (8) and (9)�are strong enough so as
never to be completely overturned by international trade costs in all �nal sectors.17
16All the results in this section can easily be extended to a general Cobb-Douglas utility function with constant
(but non-equal) expenditure shares across goods. The speci�c choice of (1) is just for algebraic simplicity.17Assumption 2 is a su¢ cient condition (but is not a necessary condition) to ensure that positive trade between
H and F always takes place in equilibrium. Intuitively, Assumption 1 creates another source of comparative
advantage in our model, in addition to heterogeneities in sectoral labor productivities. As a result, even when
" � � , our model may still deliver positive trade, provided � is su¢ ciently large.
13
From the trade balance equilibrium condition (16) we can obtain our �rst result concerning
the equilibrium relative wage, !�.
Proposition 1 In equilibrium, the wage in H is strictly greater than in F . That is, !� > 1.
Furthermore, !� is strictly increasing in �, and !� < minn(1 + �) �
12 ; (1 + ") (1 + �)�1
oif
(1 + ")2 > �, whilst !� < (1 + �)�1 �12 if (1 + ")2 � �.
The result !� > 1 is a straightforward implication of the fact that Assumption 1 conveys an
aggregate advantage by H over F . As a result, in equilibrium, ! must rise above one, in order
to allow F to be able to export to H as much as H exports to F . Notice that since labor is
the only non-reproducible input in our model, wages are also equal to income per head in each
country. Thus, Proposition 1 is ultimately stating that H is richer than F .
For future reference it proves convenient to de�ne four di¤erent thresholds for �j, namely:
�H � ln(1 + ")� ln(1 + �)� ln(!�)2 ln(1 + ")� ln(�) (17)
�H � ln(1 + ")� ln(1 + �) + ln(�)� ln(!�)2 ln(1 + ") + ln(�)
(18)
�F � ln(1 + ") + ln(1 + �)� ln(!�)2 ln(1 + ")� ln(�) (19)
�F � ln(1 + ") + ln(1 + �) + ln(�)� ln(!�)2 ln(1 + ") + ln(�)
: (20)
The above thresholds are obtained from the expressions in (11) and (12) in the following
way: �H solves P Fj;F (�H) = P Fj;H(�H) and �F solves PHj;H(�F ) = PHj;F (�F ) when �j � 0:5,
whereas �H solves P Fj;F (�H) = P Fj;H(�H) and �F solves PHj;H(�F ) = PHj;F (�F ) when �j � 0:5.
Hence, the thresholds �H and �H (resp. �F and �F ) pin down the �nal goods such that, given
the value of !�, their market price when sold in F (resp. when sold in H) would be identical
regardless of where it was originally produced. Notice that � > 0 implies �H < �F , while
Assumption 2 together with the equilibrium result !� > 1 means that �F < 1. Furthermore,
�H < �F when (1 + ")2 > �, while �H > �F holds when (1 + ")
2 < �. In addition, the results
in Proposition 1 concerning the bounds for !� imply that �H > 0 always holds.18
By using the thresholds (17)-(20), we can fully split the space of �nal goods according to
their price in the destination country, given the country of origin of the good.18The comparisons of �H vis-a-vis �H and �F vis-a-vis �F are somewhat more convoluted, as they involve sev-
eral possible combinations of parametric con�gurations and feasible solutions for !� given those con�gurations.
For example, whenever (1 + ")2 < � holds true, for any feasible values of !�, we have 0 < �H < 0:5 < �H < 1
and �F < 0:5 < �F < 1. Instead, when (1 + ")2> �, we have �H < �H i¤ !� > (1 + �)�1�
12 , and �F < �F i¤
!� > [(1 + �)�]12 holds in that range; both conditions fail to hold true for � su¢ ciently close to zero.
14
Lemma 2 From (11) and (12), and the equilibrium relative wage !�, by using (17)-(20), we
can derive the following set of conditions for P Fj;F relative to PFj;H and for P
Hj;H relative to P
Hj,F :
1. Suppose � < (1 + ")2, then:
� P Fj;H < P Fj;F for 0 � �j < ��H , while P Fj;H > P Fj;F for ��H < �j � 1, where ��H = �H if!� � (1 + �)�1 � 12 holds true, while ��H = �H if instead !� < (1 + �)
�1 �12 :
� PHj;H < PHj;F for 0 � �j < �F , while PHj;H > PHj;F for �F < �j � 1.
2. Suppose � > (1 + ")2, then:
� P Fj;H < P Fj;F for �H < �j < �H , while P Fj;H > P Fj;F for 0 � �j < �H and for
�H < �j � 1.
� PHj;H < PHj;F for maxf0; �Fg < �j < �F , while PHj;H > PHj;F for �F < �j � 1 and
0 � �j < maxf0; �Fg, where �F > 0 if and only if !� > (1 + �) (1 + ") holds true.
In equilibrium, consumers in both H and F will always buy good j from the producer who
can sell it in each market at the lower price. Hence, relying on Lemma 2, we can next derive
the equilibrium patterns of trade and specialization in the two-country world economy.
Proposition 2 The patterns of specialization and trade di¤er qualitatively depending on whether
� > (1 + ")2 or � < (1 + ")2.
i) Ricardian-based specialization: When � < (1 + ")2, trade patterns and specialization
are governed by heterogeneities in labor productivities. Country H becomes an exporter of
�nal goods whose �j 2 [0; ��H), where ��H = �H (resp. ��H = �H) if !� � (1 + �)�1 �
12
(resp. !� < (1 + �)�1 �12 ) holds true. Country F becomes an exporter of the �nal goods whose
�j 2 (�F ; 1]. Final goods whose �j 2 [��H ; �F ] are sourced locally by both H and F .
ii) Transport cost-based specialization: When � > (1 + ")2, trade patterns and specializa-
tion are governed by road network length di¤erences between H and F . Country H becomes
an exporter of �nal goods whose �j 2 (�H ; �H). Country F becomes an exporter of �nal good
whose �j 2 [0; �F ) [ (�F ; 1] if !� > (1 + ")(1 + �) holds true, while it becomes an exporter of�nal goods whose �j 2 (�F ; 1] if instead !� � (1+")(1+�) holds true. When !� > (1+")(1+�)�nal goods whose �j 2 [�F ; �H ] [ [�H ; �F ] are sourced locally by both H and F , while when
!� < (1 + ")(1 + �) this happens for those goods whose �j 2 [0; �H ] [ [�H ; �F ].
15
The patterns of trade and specialization described by Proposition 2 are graphically depicted
in Figure 1.19 The upper panel plots case i) of Proposition 2 �i.e., � < (1 + ")2�, while and
the lower panel shows case ii) �i.e., � > (1 + ")2.20 The horizontal axis of Figure 1 orders �nal
goods according to their speci�c �j 2 [0; 1]; the vertical one measures the relative wage !.Consider �rst the upper panel of Figure 1. Given a certain level of !, all the goods that
lie on the left of the solid line would be exported by H, while all the goods lying on the right
of the dashed line would be exported by F . The gap in between the two curves represents the
set of goods that would not be traded internationally. As it can be observed, the set of goods
exported by H gets smaller as ! increases. Conversely, the set of goods exported by F expands
with !. At the extremes, when ! � 1=(1 + �)(1 + ") all �nal goods would be produced in H,whereas for ! � (1+ �)(1+ ") they would all be produced in F (naturally, such extreme valuesof ! could not possibly hold in equilibrium).
At the equilibrium wage, !�, �nal goods with �j > �F are exported by F , and those with
�j < �H are exported by H.21 Hence, H becomes an exporter of the �nal goods that use input
0 more intensively (i.e., low-�j goods), while F an exporter of those which use input 1 more
intensively (i.e., high-�j goods). Intuitively, the condition � < (1 + ")2 means that di¤erences in
road network lengths between H and F are small relatively to their heterogeneities in sectoral
labor productivities. As a result, the labor productivity di¤erentials in the intermediate sectors
�dictated by (8) and (9)�become the leading source of comparative advantage, regulating trade
�ows in the model.
Consider now the lower panel of Figure 1. For values of �j > 0:5, this graph exhibits the
same qualitative features as the one in the upper panel. In fact, the interpretation of the curves
within the range 0:5 < �j < 1 is analogous in both graphs: given a !, the �nal goods that
lie on the left of the solid line would be exported by H and those lying on the right of the
dashed line would be exported by F . The main visual di¤erences between the graphs arise
19The solid line in Figure 1 is obtained by plotting �H and �H , as given by (17) and (18) but replacing the
speci�c equilibrium wage !� by a generic ! � 0. Similarly, the dashed line is obtained by plotting �F and �F ,as given by (19) and (20) for a generic ! � 0. Note that only the parts of the solid and dashed lines below
�j = 0:5 follow the expressions in (17) and (19), while only the parts above �j = 0:5 follow (18) and (20).
20For brevity, the upper panel of Figure 1 shows the sub-case where !� > (1 + �)�1�12 �implying that H
exports goods with �j 2 [0; �H)�, while its lower panel shows the sub-case where !� > (1+ ")(1+ �) �entailingthat F exports goods with �j 2 [0; �F ) and with �j 2 (�F ; 1]�. In Appendix A, Figure 1 (bis) plots the othertwo sub-cases encompassed by Proposition 2.21Notice that given the utility function (1), it must be that in equilibrium (1 � �F ) � !� = �H , where
(1� �F )� !� equals total imports by H and �H equals total exports by H.
16
when 0 < �j < 0:5. Within this range, the �nal goods located on the right of the solid line
would be exported by H, whereas those located on the left of the dashed line would be exported
by F .22 In turn, this case leads to a pattern of specialization that di¤ers quite drastically from
that one shown in the upper panel of Figure 1. When, � > (1 + ")2 ; we can observe that F
ends up exporting the �nal goods located at the two (opposite) ends of the unit set �namely,
�j 2 [0; �F ) and �j 2 (�F ; 1]�, while H becomes an exporter of those in the intermediate range
of �j �namely, �j 2 (�H ; �F )�The pattern of specialization depicted in the lower panel of Figure 1 represents the main
insight of the model. The intuition for the result rests on the fact that when � > (1 + ")2
the gap in the road network length is large relative to the heterogeneities in sectoral labor
productivities, and thus becomes the leading determinant of comparative advantages. Final
goods with intermediate values of �j require the use of both inputs in similar intensity. Given
the geographically di¤used distribution of input locations, this means that a large share of their
inputs will necessarily have to be transported along the road network. Instead, �rms producing
�nal goods with high and low values of �j can source a relatively large share of their inputs
from the same location where they are located, thus with less need to rely for it on the internal
transport network so heavily. In other words, sectors in the intermediate range of �j require
stronger use of the internal transport network than those whose �j lies on the upper and lower
spectrum of the unit interval. Accordingly, when H has a much denser road network than F ,
the former ends up specializing in the �nal goods with intermediate values of �j, and the latter
in those with more extreme values of �j.
22Analogously to the upper panel of Figure 1, as ! increases, the set of goods exported by H shrinks and
that one exported by F expands. The gap in between the curves represents the set of goods that are not traded
internationally. Finally, for ! < 1=(1 + �)(1 + ") all �nal goods would end up being produced by H, while for
! > (1 + �)�0:5 they would all end up being produced by F .
17
Figure 1: Patterns of Trade and Specialization
18
4 Empirical Predictions: From the Theory to the Data
In this section, we �rst describe how we attempt to bring to the data the main variables of
interest present in the model. Next, we explain how we approach the data on trade �ows to seek
for evidence consistent with the main novel predictions of the model. Table A.1 in Appendix
B provides some summary statistics for the main variables described below.
4.1 Main Variables of Interest
Input Narrowness
The �rst task is coming up with a measure of the breadth of the set of intermediate inputs used
by each industry. The model is quite stylized to allow a direct match between its technological
environment and real world data on inputs and outputs by sectors. In particular, in the real
world production functions tend to use more than only two intermediate goods. Furthermore,
the distinction between intermediate and �nal goods is not so clear-cut as assumed by the model,
as many goods satisfy both roles. Despite these shortcomings, we can still use the model as a
guide to construct measures of narrowness of the intermediate inputs base for industries.
In the model, sector j allocates a fraction 1 � �j of their total spending in intermediategoods on input 0 and the remainder �j on input 1. This means that sectors with very low or
very high values of �j source most of their inputs from only one intermediate sector, and thus
exhibit a narrow intermediate input base. Conversely, sectors with values of �j around one
half rely quite heavily on both inputs, and thus display a wide intermediate input base.
We formally de�ne the narrowness of the input base of sector j by the Gini coe¢ cient of
their expenditure shares across both inputs (Ginij). The greater the value of Ginij is, the
narrower input base of sector j.23 By using the fact that expenditure shares on input 0 and 1
are given, respectively, by 1� �j and �j, we can observe that:
Ginij =
(12� �j if 0 � �j � 1
2
�j � 12
if 12� �j � 1:
(21)
23Imbs and Wacziarg (2003) have previously used the Gini coe¢ cient to measure the degree of concentration
of labor and value added across di¤erent sectors. We apply the same methodology, but we use it to measure
the degree of narrowness/concentration of the intermediate input base of di¤erent industries. There are other
measures that could alternatively be used to capture the same concept; e.g., coe¢ cient of variation, log-variance,
Her�ndahl index. We also use those alternative measures in Section 4.3 as robustness checks. Clague (1991a,
1991b) have previously used the Her�ndahl index for the input concentration to measure what he called the
degree of �self-containment�of industries; this concept is essentially the same as our notion of �input narrowness�.
19
Hence, Ginij = 0 when �j = 0:5, while it grows symmetrically as �j moves away from its
central value of 0:5 towards either extremes on 0 and 1.
To construct a measure of input narrowness analogous to that one in (21), but based on the
available real world data, we resort to the input-output (IO) matrix of the US in 2007 from
the Bureau of Economic Analysis (BEA).24 The IO matrix comprises 389 sectors/industries.
Although the IO matrix exhibits the same number of sectors producing intermediate goods
as those producing �nal output, we restrict the set of �nal goods to those also present in the
international trade data (see description below). Thus, we index by k = 1; 2; :::; K each of the
sectors present in the IO matrix and also in the trade data, and by n = 1; 2; :::; N each of the
sectors selling intermediate inputs.
We let Xk;n � 0 denote the total value of intermediate good n purchased by sector k.
De�ning Sk;n � Xk;n=PN
n=1Xk;n � 0 as the share of n over the total value of intermediates
purchased by k, we can compute the Gini coe¢ cients analogously to those in (21). Namely,
Ginik =2�
PNn=1 n� Sk;n
N �PN
n=1 Sk;n� N + 1
N; (22)
where the argumentPN
n=1 n� Sk;l in the numerator of Ginik is ordering intermediates in non-decreasing order (i.e., Sk;n � Sk;n+1).In the empirical analysis in Section 4.3, we use Ginik to measure the degree of narrowness of
the input base of sector k. Large values of Ginik are the result of sector k sourcing most of their
intermediate inputs from relatively few sectors. Conversely, small values of Ginik tend to occur
when the distribution of Sk;n is quite evenly spread across a large number of intermediates.25
Notice �nally the link between Ginik in (22) and Ginij in (21): the former boils down to the
latter when N = 2, and Sk;n = �k;n with �k;1 + �k;2 = 1.
Export Specialization
In order to measure the degree of export specialization by sectors we use the data on trade �ows
from COMTRADE compiled by Gaulier and Zignago (2010). We use only trade �ows in year
2014. The data are categorized following the Harmonized System (HS) 6-digit classi�cation,
with 5,192 products. We map the trade �ows data based on the HS 6-digit classi�cation to the
BEA industry codes using the concordance table between the 2002 IO matrix commodity codes24This data is publicly available from https://www.bea.gov/industry/io_annual.htm.25In the extreme (unequal) case in which Sk;n0 = 1 for some n0 and Sk;n = 0 for all n 6= n0, (22) yields
Ginik = 2� [(N + 1)=N ], which approaches 1 as N !1. On the other hand, in the case when Sk;n = Sk > 0for all n = 1; :::; N , we would have Ginik = 0.
20
and the HS 10-digit classi�cation from the BEA website (after grouping the HS 10-digit codes
into HS 6-digit products). In the cases in which an HS-6 product maps into more than one
BEA code, we assign their trade �ows proportionally to each of the BEA sectors in which it
maps.26 Lastly, the IO industry codes of the 2002 classi�cation are matched to those of the 2007
classi�cation, which are the ones actually used in the computation of the Gini coe¢ cients.27
Road Network
The last key variable in our model is the length of the road network of country c (rc). We
take the road network length by countries from the data on roadways from the CIA World
Factbook. Roadways are de�ned as �total length of the road network, including paved and
unpaved portions�. The year of the data point for each country varies, ranging from year 2000
to 2016, with the median year of the sample being 2010. (See details of this variable in Table
A.8 in Appendix C.) When de�ning our empirical counterpart of the variable rc, we divide the
length of the road network by the total area of the country: rc � roadwaysc=areac. In someof the robustness checks, we use also two additional measures of transport density: waterways
density (de�ned as waterwaysc=areac) and railway density (de�ned as railwaysc=areac). The
data on length of waterways and railways are also taken from the CIA World Factbook.
4.2 Road Density and Patterns of Specialization: Testing the pre-
dictions of the model
The two-country model presented in Section 3 predicts that when heterogeneities in road net-
works are su¢ ciently large, the patterns of specialization and trade �ows follow those depicted
by the lower panel of Figure 1. More formally, when the condition � > (1 + ")2 holds true,
the country with the longer road network (i.e., country H) will export goods with intermediate
26There are 526 HS-6 products that map into two BEA Input-Output industry codes, 96 products that map
into three IO codes, 33 products that map into four IO codes, and 11 products that map into �ve or more IO
codes. (We excluded the 11 products that map into �ve or more IO codes.) None of the regression results in
Section 4.3 are signi�cantly altered when all the HS-6 products that map into more than one BEA Input-Output
industry code are dropped from the sample.27Unfortunately, we are not aware of any correspondence table between BEA 2007 codes and the HS codes,
hence we indirectly link them via the BEA 2002 codes. In the end, after mapping the HS 6-digit products into
the BEA 2002 codes, and mapping the BEA 2007 codes to the 2002 codes, we are left with data on trade �ows
and input narrowness for 294 industries as coded by the BEA 2002 classi�cation. Of the original 389 codes,
only 307 are matched to HS 10-digit codes. We have export data for 303 industries among those 307. Nine
other industries are lost when matching the BEA 2007 codes to those in BEA 2002.
21
values of �j, while the country with the shorter road network (i.e., country F ) will export
goods with values of �j located on the extremes of the unit continuum. Conceptually, this
prediction can be interpreted as stating that countries with denser road networks will tend to
exhibit a comparative advantage in the types of goods that require a wider (or more diverse)
set of intermediate inputs.
From an empirical viewpoint, if road network length di¤erences across countries shaped
somehow their patterns of specialization as our model predicts, we should then observe the
following: economies with a greater rc will tend to export relatively more of the goods produced
in industries with a smaller value of Ginik vis-a-vis economies with smaller rc. We test this
prediction using the following regression:
ln(Expoc;k) = � � (rc �Ginik) + � ��k;c + &c + �k + �c;k; (23)
In the regression equation (23) the dependent variable is given by the natural logarithm of
the total value of exports in industry k by country c to all other countries in the world in year
2014. The term (rc�Ginik) interacts the measure of input narrowness de�ned in (22) with themeasure of road density (i.e., length of roadways per square kilometer). �k;c denotes a vector
of additional covariates that may possibly in�uence specialization across countries in industries
di¤ering in terms of the degree of input narrowness. &c and �k denote country �xed e¤ects and
industry �xed e¤ects, respectively, and �c;k represents an error term.
The main coe¢ cient of interest in (23) is �. If, as the model predicts, countries with a
denser road network (i.e., countries with a greater rc) indeed tend to exhibit a comparative
advantage in goods from industries that require a wider set intermediate inputs (i.e., industries
with a smaller Ginik), then the data should deliver a negative estimate of �.
4.3 Empirical Results
Table I displays the �rst set of estimation results corresponding to (23). Column (1) includes
only our main variable of interest (i.e., rc � Ginik), together with the exporter and industrydummies. The correlation is negative and highly signi�cant, suggesting that countries with
denser road networks tend to export relatively more of the �nal goods whose production process
requires a wider intermediate input base (i.e., those exhibiting a lower Ginik).
Columns (2) -(4) in Table I show the results of this simple correlation when input narrowness
is measured by three alternative measures: the Her�ndahl index, the coe¢ cient of variation,
and the log-variance of industry k�s intermediates expenditure shares (Sk;n). The estimate of
� is negative and highly signi�cant under all these alternative measures as well. Finally, in
22
(1) (2) (3) (4) (5)
Road Density x Input Narrowness 4.084*** 0.710*** 0.034*** 0.122*** 3.960***(0.305) (0.141) (0.004) (0.012) (0.814)
Road Density 4.060***(0.837)
Observations 42,578 42,578 42,578 42,578 42,578Rsquared 0.765 0.764 0.764 0.764 0.263Country FE Yes Yes Yes Yes NoSector FE Yes Yes Yes Yes YesDependent Variable log(Expo c,k ) log(Expo c,k ) log(Expo c,k ) log(Expo c,k ) log(Expo c,k /Expo c )Number of Countries 166 166 166 166 166Narrowness Measure Gini Herf Coef Var Log Var GiniRobust standard errors reported in parentheses in colums (1) (4), and clustered at the country level in column (5). The dependent variable in columns (1) (4) is the log of
total exports in industry k by country c in year 2014. The dependent variable in column (5) is the log of the share of exports in industry k by country c over total exports
by country c in year 2014. The number of industries is 294 in all regressions. *** p<0.01, ** p<0.05, * p<0.1
TABLE IExport Specialization across Industries with Different Levels of Input Narrowness
column (5) displays the result of a regression that includes the length of the road network as a
regressor, together with the interaction term rc �Ginik.28 Since Ginik is always smaller thanunity, the results show that countries with longer road networks tend to export more across the
board, but the increase in exports is more pronounced in industries whose Ginik is smaller.
In Table II we sequentially incorporate some additional interaction terms that may also in-
�uence the patterns of specialization across industries with di¤erent levels of input narrowness.
Column (1) adds an interaction term between Ginik and an index of Rule of Law, taken from
World Governance Indicators. The rationale behind including this term lies on the argument
in Levchenko (2007) and Nunn (2007), who show that countries with better contract enforce-
ment institutions display a comparative advantage in industries that are heavily dependent on
relationship-speci�c investments. Within our context, industries that need to source a broader
set of intermediates may bene�t relatively more from a sound legal environment, as they need
to establish relationships with a greater number of input providers. Given that countries with
better institutions tend to be also richer and invest more in basic infrastructure, omitting this
term could lead to an overestimation (in absolute value) of the correlation coe¢ cient of inter-
est in (23). The regression in column (1) of Table II yields indeed a negative and signi�cant
coe¢ cient associated with the interaction term between rule of law and Ginik, consistent with
28Since in column (5) we have to drop the country �xed e¤ects &c from the regression, there we use as
dependent variable the log of the export share in industry k by country c �i.e., ln(Expoc;k=Expoc)�rather than
simply ln(Expoc;k), so as to take care of the size of exports by each country. The estimates in column (5) remain
essentially the same if alternatively using ln(Expoc;k=GDPc) as dependent variable.
23
the previous literature on institutions and specialization. In addition, while the magnitude ofb� falls relative to column (1) of Table I, it still remains negative and signi�cant.Another possible source of omitted variable bias is related to the e¤ect of �nancial markets.
There is a large body of literature that sustains that �nancial markets are instrumental to
opening new sectors and increasing the variety of industries in the economy (e.g., Greenwood
and Jovanovic, 1990; Saint-Paul, 1992; Acemoglu and Zilibotti, 1997). We could then expect
that countries with more developed �nancial markets would also be better able to specialize in
industries that require a wider input base. To deal with this concern, in column (2) we interact
the Gini coe¢ cients with an indicator of �nancial development: the log ratio of private credit to
GDP. (This indicator is taken from the World Bank Indicators database, and averaged during
years 2005-2014.) As we can readily observe, the e¤ect of �nancial development interacted
with Ginik is signi�cant and it carries a sign consistent with the past literature on growth and
diversi�cation. Yet, the estimate of � still remains negative and signi�cant.
Column (3) adds an interaction term between Ginik and log GDP per capita. This term
would control for the possibility that richer economies may be better able to produce goods
with lower Ginik, as in general richer economies tend to exhibit a more diversi�ed productive
structure than poorer ones. As we can see, the results concerning � remain essentially unaltered.
In column (4) we introduce two additional regressors to control for specialization driven
by factor endowments: i) an interaction term between capital intensity of industry k and the
(log) stock of physical capital per worker in country c; ii) an interaction term between the skill
intensity of industry k and the (log) stock of human capital in country c. The measures of
capital and skill intensity at the industry level are constructed from the NBER-CES Manufac-
turing Industry database, for year 2011.29 Some industries are lost from the sample when we
introduce the industry factor intensity measures, since the NBER dataset contains information
only for manufacturing industries. For comparability, in column (5) we display the results of
the regression in column (3), but using the restricted sample. The coe¢ cients associated to
the factor intensities carry the expected sign, while the estimates of � remain negative and
signi�cant. Furthermore, the estimated coe¢ cients are of similar magnitude in both columns.
Finally, the last two columns of Table II address the possibility of a di¤erential e¤ect of
the road network on the pattern of specialization depending on the population density of the29Capital intensity is computed as the total stock of physical capital per worker by industry. Skill intensity is
measured by the average wage by industry. (See Becker, Gray and Marvakov (2013) for details on the NBER-
CES Manufacturing Industry database.) Both the measure of physical capital per worker and the index of human
capital are drawn from the Penn Tables database. (The human capital index is based on the average years of
schooling from Barro & Lee (2013) and an assumed rate of return of education based on Mincer estimates.)
24
(1) (2) (3) (4) (5) (6) (7)
Road Density x Input Narrowness 2.240*** 2.382*** 2.356*** 1.852*** 1.845*** 2.956*** 2.439***(0.334) (0.363) (0.364) (0.379) (0.385) (0.437) (0.453)
Rule of Law x Input Narrowness 5.414*** 2.639*** 2.435*** 3.132*** 3.045*** 2.304*** 2.999***(0.441) (0.648) (0.702) (0.762) (0.766) (0.703) (0.765)
log (Priv Credit/GDP) x Input Narrowness 3.743*** 3.547*** 1.704* 1.681* 3.224*** 1.500*(0.760) (0.822) (0.913) (0.915) (0.828) (0.918)
log GDP per capita x Input Narrowness 0.417 0.905 1.306** 0.522 0.822(0.602) (0.656) (0.654) (0.604) (0.658)
Capital Intensity x log (K /L )c 0.010** 0.010**(0.004) (0.004)
Skill Intensity x log H c 0.008*** 0.008***(0.001) (0.001)
(Pop) Density x Input Narrowness 0.333*** 0.197**(0.107) (0.101)
Road Dens x (Pop) Dens x Input Nwness 0.089*** 0.064**(0.028) (0.027)
Observations 41,947 40,692 40,692 31,892 31,892 40,692 31,892Rsquared 0.764 0.764 0.764 0.794 0.793 0.764 0.794Number of Countries 163 157 157 134 134 134 157Number of Industries 294 294 294 259 259 294 259Robust standard errors in parentheses. All regressions include country and industry fixed effects. The dependent variable is the log of exports in industry k by country c in 2014.Rule of law is taken from the World Governance Indicators (WB) for year 2014. Private credit over GDP is taken from the World Bank Indicators, averaged for years 20052014.GDP per capita, stock of physical capital, and the human capital index are taken from the Penn Tables, all for 2014. Measures of physical capital and skill intensity by industryare taken from the NBERCES Manufacturing Industry Database, and corresponds to year 2011. *** p<0.01, ** p<0.05, * p<0.1
TABLE IIExport Specialization across Industries with Different Levels of Input Narrowness: Additional Covariates
25
economy. One could expect that more densely populated countries may display also a greater
concentration of activities in fewer locations. Hence, all else equal, more densely populated
countries may need to resort less strongly on a vast road network than sparsely populated
countries do. Columns (7) and (8) assess this possibility by introducing an interaction term
between population density and Ginik, and a triple interaction term which also includes rc.
If road network length is especially important for specialization in economies that are less
densely populated, then the triple interaction term should carry a positive estimate. As can
be observed, this is indeed the case. Moreover, the estimate of � after introducing the triple
interaction term is still negative and highly signi�cant.
Table III displays some of the regressions previously presented in Table II, but now splitting
the sample of countries in two subsamples, according to whether their income is above or below
the median. The odd-numbered columns show the results for the subsample of �high-income
countries�, while the even-numbered columns do that for the �low-income countries�. The results
show that the e¤ect of road density on pattern of specialization holds true both for richer and
poorer countries. In addition to that, the e¤ect seems to be consistently greater in magnitude
for the subsample of economies whose income is below the median.30
Additional Robustness checks
Some further robustness checks are provided in Appendix B in Tables A.2, A.3, and A.4.
Tables A.2 and A.3 change the measure of transport density used in the previous regressions.
In Table A.2, we show the results of a set of regressions substituting rc in (23) by railway density,
computed as the total railway network length of country c per square kilometer. In Table A.3,
we expand our measure transport network length to include (in addition to roadways) also the
total length of internal railways and waterways. All the results in Table A.2 and A.3 follow a
similar pattern as those previously shown in Table I and II.
Table A.4 shows that all the previous results are also robust to: i) excluding very small
countries (both in terms of area and population), ii) excluding very large countries (in terms
of area), iii) controlling for the e¤ect of area and population (in both cases interacted with the
measure of input narrowness), iv) including the interaction between total GDP and input nar-
rowness, and v) excluding from the sample those countries whose road networks were measured
before year 2010 (which is the median year in the sample).
The rationale for these additional robustness checks is the following. In the case of very
30This di¤erence in magnitude could suggest the presence of some sort of decreasing marginal e¤ect of road
density, since richer economies tend to exhibit denser road networks than poorer ones (see Figure 2 later on).
26
(1) (2) (3) (4) (5) (6)
Road Density x Input Narrowness 2.118*** 4.935*** 1.636*** 4.495*** 1.635*** 4.477***(0.354) (1.482) (0.373) (1.712) (0.375) (1.713)
Rule of Law x Input Narrowness 0.673 4.075** 1.393 2.865 1.170 2.810(1.005) (1.739) (1.094) (2.025) (1.091) (2.029)
log (Priv Cred/GDP) x Input Narrowness 4.882*** 2.001 4.640*** 0.758 4.712*** 0.739(1.086) (1.319) (1.193) (1.428) (1.190) (1.428)
log GDP per capita x Input Narrowness 0.418 1.723 2.002 2.087* 2.357 2.459**(1.541) (1.069) (1.638) (1.112) (1.650) (1.111)
Capital Intensity x log (K /L )c 0.010 0.014*(0.012) (0.009)
Skill Intensity x log H c 0.014*** 0.015***(0.002) (0.002)
Observations 22,087 18,605 17,137 14,755 17,137 14,755Rsquared 0.791 0.632 0.808 0.652 0.807 0.651Countries Sample (high/low income) High Low High Low High LowNumber of Countries 79 78 68 66 68 66Number of Industries 294 294 259 259 259 259Robust standard errors reported in parentheses. All regressions include country and industry fixed effects. The dependent variable is log (Expo c,k ) in the year 2014.The highincome sample comprises countries with GDP per capita above the sample median, and the lowincome sample the countries with GDP per capita below it.The median income of the sample lies between that of Iraq ($12,095 PPP) and South Africa ($12,128 PPP) in 2014. Rule of law is taken from World GovernanceIndicators for year 2014. Private credit over GDP is taken from the World Bank Indicators, averaged for years 20052014. Physical capital and skill intensity byindustry are taken from the NBERCES Manufacturing Industry Database, and corresponds to year 2011. *** p<0.01, ** p<0.05, * p<0.1
TABLE IIIHighIncome and LowIncome Subsamples
small countries, on the one hand, they may �nd it easier to link together geographic locations
while, on the other hand, they may be less able to provide su¢ cient opportunities for input
diversity. Next, regarding very large countries in terms of area, the concern could be that some
of those countries may contain very large portions of uninhabitable land, which may end up
turning our measure of road density somewhat imprecise in those cases. Controlling for the size
of the country (both in terms of area and population) takes into account the possibility that
larger countries may face more opportunities for input diversity, regardless of the density of
their internal transport network. Similarly, including total GDP can control for the possibility
that there exist minimum size requirements to open up some sectors in the economy. Finally,
restricting the sample to countries whose road networks were measured after 2010 helps in
harmonizing the data year on trade �ows and road density, and shows that the found e¤ects
are not contaminated by countries whose data on road networks is relatively older.
27
5 Endogeneity and Alternative Interpretations
The previous section presented a robust association between road density in country c and its
degree of specialization in industries that rely on a wide set of inputs. While those results
are certainly consistent with the main predictions of the model, they cannot be taken as hard
evidence of its underlying mechanism. Two separate issues deserve further discussion and
analysis. First, the correlation found in the previous regressions could as well be the result of
road infrastructure responding to transport needs stemming from industry specialization (i.e.,
reverse causation). Second, our interpretation of a lower value of Ginik as re�ecting greater
need of industry k for the local transport infrastructure is debatable, as previous authors have
looked at that variable as capturing a di¤erent feature: the degree of product complexity of
industry k. In the next two subsections we aim to address these two points more explicitly.
5.1 Endogeneity and Reverse Causation
Our model has resorted to two critical assumptions that warrant further discussion in case the
previous empirical results are intended to be interpreted as evidence of a causal e¤ect from road
density to specialization. Firstly, it has taken rc as exogenously given. The length of a country�s
road network is however the result of investment choices in infrastructure, and hence it will
respond to a host of economic variables and incentives. Secondly, the model has assumed away
any sort of intrinsic di¤erences in productivities directly linked to the production functions of
�nal goods. In fact, all di¤erences in countries�productivities across �nal sectors arise indirectly
from the heterogeneities in the intensity of inputs implied by the parameter �j in (3).
Relaxing the two above-mentioned assumptions can easily lead to a model where � in
(23) can be confounding an e¤ect from road density to specialization, together with reverse
causality from the latter to the former. For example, suppose that for some reason the �nal
good production functions di¤er across countries, and that H is relatively more productive than
F in the �nal sectors whose �j lies near one half.31 In a context like this one, if countries can
invest in expanding their road networks, we could well expect rH to be larger than rF simply
because the incentives to do so are greater in H than in F . From an empirical viewpoint, this
31For example, we may have that �nal good productions functions are given by (3) for �j 2 [0; 0:5� �] andfor �j 2 [0:5 + �; 1], where 0 < � < 0:5, and by
Yj = (1 + �c)1
��jj (1� �j)1��j
X1��j0;j X
�j1;j ; for �j 2 (0:5� �; 0:5 + �); with �H > �F :
28
reasoning means that � could end up capturing (at least partially) an e¤ect going from patterns
of specialization to road density.
The rest of this subsection provides some further support for the notion that the density of
the internal transport network is instrumental to specialization in industries with wider input
bases. First, we show that a similar correlation to that one found in Section 4.3 arises when the
density of the transport network is measured by the density of internal waterways. Next, we
show that the results in Section 4.3 remain true when we instrument the density of a country�s
road network with topographical measures of terrain roughness.
5.1.1 Patterns of Specialization and Waterways Density
This �rst part intends to provide some further evidence consistent with the main mechanism
of the model, by relying on a measure of countries�transport network that is less sensitive to
reverse causality concerns than rc is. We measure now the internal transport network of an
economy by the density of their waterways network. We draw the data on waterways from
the CIA World Factbook, and de�ne waterways density as waterways length per square km.32
Arguably, while countries can still expand their waterways by investing in creating canals or
improving the navigability of some rivers and bodies of water, the scope for this is far more
limited than in the case of roads.
One additional aspect we investigate here is the possibility that waterways impact special-
ization heterogeneously at di¤erent stages of development. For various reasons, richer economies
tend to have much denser road networks than poorer ones. In particular, poorer economies may
�nd it harder to undertake the necessary investment to build a su¢ ciently developed road in-
frastructure. On the other hand, while the presence of waterways may have in�uenced patterns
of development before railroads and roads became more widespread worldwide, waterways are
no longer a mode of transportation that seems to be associated with economies�current level
of development. In fact, a quick look at simple cross-country correlations in Figure 2 shows
that income per head and road density display a clear positive correlation, while the association
between income per head and waterways density is rather weak.33
32The CIA World Factbook measures waterways as the total length of navigable rivers, canals and other
inland bodies of water.33One possible interpretation of the correlations in Figure 2 is that, as economies grow richer, roadways tend
to gradually overshadow waterways as a mode of internal transportation. From this perspective, we could then
expect waterways to represent an important determinant of patterns of specialization in poorer economies, but
losing preeminence in richer economies where roadways can more easily make up for an insu¢ ciently dense
internal waterway network.
29
Figure 2: Roadways and waterways density against GDP per head
Table IV displays the results of a regression equation analogous to (23), but where rcis replaced by a measure of waterways density. The table shows the results of two sets of
regressions for three di¤erent countries sample: entire sample, high-income countries, and low-
income countries.
The regressions based on the whole set of countries tend to yield an estimate that is nega-
tive. However, this aggregate result masks important heterogeneities in the e¤ect of waterways
density on export specialization in the case of richer versus poorer economies. Column (2)
shows that waterways density carries no impact at all in the subsample of above-median in-
come economies. By contrast, column (3) exhibits a negative and highly signi�cant coe¢ cient.
This suggests that, in the case of poorer economies, those that enjoy a denser network of wa-
terways tend to export relatively more in industries that require a wider intermediate input
base. Columns (4)-(6) re-run the regressions in columns (1)-(3) but also including the original
interaction term rc �Ginik. The results for the impact of waterways density on specializationin (5) and (6) follow a very similar pattern as those in (2) and (3). Moreover, the regressions
also show that the coe¢ cient for road density remains negative and signi�cant.34
Finally, Table A.5 in Appendix B shows the results a of set of regressions that include a
triple interaction term between rc, Ginik and log income, rather than splitting the sample of
countries according to income. All the results remain qualitatively consistent with those in
Table IV.34Note that the results in columns (5) and (6) of Table IV are not directly comparable to those in columns
(3) and (4) of Table III due to the loss of some countries in the samples of Table IV.
30
(1) (2) (3) (4) (5) (6)
Road Density x Input Narrowness 0.262* 0.016 1.120*** 0.137 0.194 1.100***(0.148) (0.140) (0.347) (0.170) (0.172) (0.347)
Rule of Law x Input Narrowness 4.205*** 4.727*** 7.389*** 3.576*** 4.257*** 4.533*(0.838) (1.546) (2.511) (0.902) (1.561) (2.811)
log (Priv Cred/GDP) x Input Narrowness 1.724* 2.403* 0.484 1.857* 2.505* 0.624(1.077) (1.377) (1.664) (1.085) (1.382) (1.663)
log GDP per capita x Input Narrowness 0.457 5.220* 0.283 0.620 5.752** 0.830(0.867) (2.944) (1.256) (0.874) (2.962) (1.268)
Capital Intensity x log (K /L )c 0.007 0.014 0.021* 0.008 0.013 0.021*(0.006) (0.014) (0.012) (0.006) (0.014) (0.012)
Skill Intensity x log H c 0.007*** 0.020*** 0.018*** 0.007*** 0.020*** 0.018***(0.001) (0.003) (0.002) (0.001) (0.003) (0.002)
Road Density x Input Narrowness 1.381** 1.426** 7.249***(0.711) (0.703) (2.386)
Observations 22,357 11,750 10,607 22,357 11,750 10,607Rsquared 0.811 0.798 0.712 0.811 0.798 0.712Countries Sample All High Low All High LowNumber of Countries 93 46 47 93 46 47Number of Industries 259 259 259 259 259 259Robust standard errors in parentheses. All regressions include country and industry fixed effects. The dependent variable is log(Expoc,k ) in year 2014. Waterways is taken from theCIA World Factbook. It comprises total length of navigable rivers, canals and other inland water bodies. Waterway density equals internal waterways per square km. Columns (2) and(5) include only those countries with above median income in the sample. Columns (3) and (6) include only those countries whose income lies below the median income in the sample. ***p<0.01, **p<0.05, *p<0.1
TABLE IVWaterways Density as Measure of Transport Network
31
5.1.2 Instrumental Variables: Terrain Roughness and Roadway Density
This second part intends to address more directly the concern of reverse causation from industry
specialization to road density. To do so we instrument rc with measures of terrain roughness in
country c. The idea is drawn from Ramcharan (2009), who shows that countries with rougher
terrain surface tend to exhibit less dense road networks.35 In the context of our paper, if the
roughness of the terrain a¤ects the density of the internal road network, but it does not exert a
systematic impact on specialization across industries with varying degrees of input narrowness
via other alternative channels, then it can serve as a valid instrument for rc.36
In Table V we show the results of the two-stage least square (2SLS) regressions using three
alternative measures of terrain roughness: i) the di¤erence between the maximum and minimum
land elevation in country c, taken from the CIA Factbook; ii) the standard deviation of elevation
of country c measured at 30�degree grids (approx. 1km cells), taken from Ramcharan (2009);
iii) the percentage of mountainous terrain, taken from Fearon and Laitin (2003). Table A.6 in
Appendix B shows that all three measures of terrain roughness are negatively correlated with
road density, even after controlling for several other country-level variables.37
35Ramcharan (2009) studies the spatial concentration of economic activities within countries, and how this
is a¤ected by their surface topography. The author argues that countries with rougher topography tend to
display stronger spatial concentration of economic activity, and that this is partly of explained by the poorer
land transportation associated with rougher terrain.36Notice that a violation of the exclusion restriction in this context requires more than simply terrain roughness
having an impact on industry specialization. For the exclusion restriction to be violated, it must be the case
that terrain roughness a¤ects specialization across industries in a way that is also correlated with their degree
of input narrowness (besides the e¤ect mediated by the impact of terrain roughness on road density). For
example, the exclusion restriction may be threatened if economies with rougher terrain tend to also display
more heterogeneous climatic conditions and land con�gurations, and this allows them to enjoy a more diverse
productive structure. Conversely, it may be that rougher terrain reduces the share of inhabitable land, curtailing
productive heterogeneity, and thereby possibly leading to a violation of the exclusion restriction via a negative
impact of roughness on specialization in industries with wide input breadth. While we cannot test the validity
of the exclusion restriction, the results in Table A.7 in Appendix B (see also the discussion therein) are in
principle encouraging about how concerned we should remain about a the possibility of a direct impact of
terrain roughness, besides that one mediated by its e¤ect on road density.37All the measures of terrain roughness used here aim at capturing large-scale terrain irregularities. In a sense,
these seem to be the types of irregularities that can most severely hinder internal transport networks. Other
papers, e.g. Nunn and Puga (2012), have resorted to the methodology developed by Riley et al. (1999) so as to
measure small-scale terrain irregularities. While small-scale terrain roughness measures seem more appropriate
for capturing the presence of small geographic formations that may provide natural sources of protection to
certain groups of people, they may not represent the main source of obstruction to dense transport networks.
32
Columns (1) and (2) display the results of the 2SLS regressions based on the di¤erence
between the maximum and minimum land elevation in c as instrument for rc. Arguably, this
variable may be seen as a relatively crude measure for terrain roughness. However, it has the
upside of being available for the exact same samples as those in Section 4.3. As a result, columns
(1) and (2) of Panel A can be directly compared to their respective OLS counterparts in columns
(3) and (4) in Table II. Next, columns (4) and (6) display the estimation results of the 2SLS
regressions in which the instrument for road density is based on the standard deviation of land
elevation. Since for some of the countries in the original sample this information is missing,
columns (3) and (5) show their respective OLS estimates for the corresponding sample. Finally,
columns (8) and (10) report the 2SLS results when using the percentage of mountainous terrain
as instrumental variable. Again, in the sake of comparability, columns (7) and (9) report the
OLS estimates for the corresponding country samples.
The main message to draw from Table V is that the 2SLS regressions consistently yield
a negative and signi�cant estimate for our coe¢ cient of interest. These results reinforce the
support for the hypothesis that the density of the internal transport network is an important
determinant of specialization in industries with wide input bases, by exploiting the variation in
the internal road network across countries predicted by their degrees of terrain roughness.
One additional point to note from Table V is that the 2SLS estimates for � tend to be
consistently greater in absolute magnitude than their OLS counterparts. This would in prin-
ciple run against the direction of the bias that would stem from the reverse causality concern
discussed in the second paragraph of Section 5.1 (i.e., the notion that economies specializing in
industries that rely on a wide variety of inputs may tend to invest more in transport infrastruc-
ture). One possible reason behind these results is that the instrument may also be alleviating
some degree of measurement error in our indicator of road density. In that respect, recall that
rc is computed using the total length of roads by country. This disregards the fact that di¤erent
roads may di¤er substantially in terms of quality and width, and it is also summing up together
paved and unpaved roads. Furthermore, the total length or the road network is not taking into
account the possibility of a very ine¢ cient lay out of the network. All these issues could end
up reducing the precision with which rc captures the notion that a denser road network allows
cheaper internal transportation of inputs. Therefore, when instrumenting rc, we may not only
be dealing with problems of endogeneity, but also with the fact that in some cases our measure
of road density may be quite imprecisely gauging the e¢ ciency of the internal road network.
33
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)2SLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS
Road Density x Gini 5.006*** 3.188* 3.364*** 6.199** 2.204*** 6.166* 3.347*** 5.638* 2.000*** 5.645*(1.513) (1.670) (0.509) (2.800) (0.544) (3.573) (0.506) (3.258) (0.515) (3.355)
Rule of Law x Gini 1.260 2.469** 2.285*** 1.306 2.886*** 1.264 2.082*** 1.108 3.147*** 1.543(0.943) (1.100) (0.720) (1.187) (0.786) (1.635) (0.737) (1.564) (0.790) (1.661)
Fin Dev x Gini 3.340*** 1.678* 4.063*** 3.978*** 2.605*** 2.849*** 3.975*** 4.060*** 1.507 1.719*(0.838) (0.917) (0.878) (0.890) (0.956) (0.963) (0.870) (0.874) (0.938) (0.948)
log GDP pc x Gini 0.088 1.077* 0.161 0.620 1.170* 1.857** 0.056 0.437 0.926 1.562*(0.623) (0.677) (0.667) (0.781) (0.714) (0.908) (0.641) (0.820) (0.676) (0.868)
K Intens x log (K /L )c 0.010** 0.009** 0.009** 0.008* 0.008*(0.004) (0.0045) (0.0045) (0.004) (0.004)
Skill Intens x log H c 0.008*** 0.007*** 0.007*** 0.008*** 0.008***(0.001) (0.001) (0.001) (0.001) (0.001)
Observations 40,692 31,892 35,988 35,988 29,229 29,229 36,544 36,544 30,067 30,067Rsquared 0.763 0.794 0.764 0.764 0.794 0.793 0.757 0.757 0.795 0.795Number of Countries 157 134 135 135 121 121 138 138 126 126Number of Industries 294 259 294 294 259 259 294 294 259 259
Terrain Roughness x Gini 0.013*** 0.013*** 0.452*** 0.392*** 0.007*** 0.007***(0.0004) (0.001) (0.017) (0.019) (0.0002) (0.0003)
Rule of Law x Gini 0.344*** 0.371*** 0.290*** 0.350*** 0.398*** 0.413***(0.011) (0.012) (0.011) (0.012) (0.011) (0.012)
Fin Dev x Gini 0.109*** 0.059*** 0.066*** 0.029*** 0.004 0.032***(0.011) (0.013) (0.011) (0.011) (0.009) (0.010)
log GDP pc x Gini 0.144*** 0.164*** 0.165*** 0.184*** 0.152*** 0.163***(0.007) (0.010) (0.009) (0.010) (0.008) (0.009)
K Intens x log (K /L )c 0.005 0.010 0.008(0.008) (0.007) (0.007)
Skill Intens x log H c 0.010 0.017 0.010(0.015) (0.013) (0.014)
FStat: PValue 0.00 0.00 0.00 0.00 0.00 0.00Robust standard errors in parentheses. All regressions include country and industry fixed effects. The dependent variable in Panel A is the log(Expoc,k ) in year 2014. 'Terrain Roughness' ismeasured as follows: i ) in columns (1) and (2) by the diffference between the maximum and the minimum elevation in country c (source: CIA World Factbook); ii ) in columns (4) and (6)by the std. deviation of elevation in country c computed at the 30'' degree resolution (source: Ramcharan, 2009); iii ) in columns (8) and (10) by the percentage of mountainous in terrainin country c (source: Fearon and Laitin, 2003). *** p<0.01, ** p<0.05, * p<0.1
TABLE VTwoStage Least Squares Regressions using Terrain Roughness as Instrument for Roadway Density
PANEL A: SECONDSTAGE RESULTS (AND OLS COMPARISONS)
PANEL B: FIRSTSTAGE RESULTS (Dep. Variable: Road Density x Input Narrowness)
34
5.2 Alternative Interpretations of the Input Breadth Measures
The analysis in Section 4 was based on the notion that the degree of input breadth of industry k
can serve as an indicator for how reliant this industry is on the internal transport network. The
need to source a large variety of inputs can certainly make a particular sector heavily dependent
on e¢ cient transportation; however, it can also mean that the sector is highly sensitive to
sound contract enforcement. Indeed, Blanchard and Kremer (1997) and Levchenko (2007) have
previously used input-output data to compute diversi�cation indices for intermediate input
purchases across industries, and use them to proxy the degree of complexity of sectors: sectors
with more diversi�ed (i.e., less concentrated) input bases are considered to be more complex.38
In their analysis, more complex sectors require better contract enforcement to work e¢ ciently.
From this viewpoint, countries with better functioning institutions should exhibit a comparative
advantage in industries with broad intermediate input bases. Levchenko (2007) shows that this
is indeed the case for US imports: the US imports a higher share of goods with greater diversity
of intermediate inputs from countries with better rule of law.
The regressions in Tables II - IV have conditioned on the interaction between rule of law
in country c and the Gini coe¢ cient for intermediate inputs in industry k. The estimate of �
remained consistently negative and signi�cant, regardless of the introduction of this additional
control. In that regard, our results seem to suggest that both institutions and local transport
networks are instrumental and complementary to the growth of industries with wide input
bases. This section will attempt to further strengthen this argument
Countries with better institutions are in general richer, and also exhibit a denser transporta-
tion infrastructure network. If Ginik�rc in (23) were not capturing any type of impact relatedto how transport-intensive industry k is, but only the e¤ect of rule of law in country c through
its correlation with rc, then the correlation found in Table I should arise more prominently for
industries that are relatively more dependent on judicial quality. The regressions reported in
Table VI show this is actually not the case in the data.
Columns (1) and (2) in Table VI show the results of the simple correlation reported initially
in column (1) of Table I, after splitting the set of industries in two subsamples: low contract
intensity vs. high contract intensity. To do so, we take the measure of contract intensity by
industries from Nunn (2007), and split the sample of industries according to whether they rank
below or above the median value of contract intensity.39 If the Ginik were simply proxying for
38Both articles used the Her�ndahl index of concentration instead of the Gini as their benchmark measure.39Nunn (2007) reports contract intensity measures for 222 industries coded according to NAICS 1997. We
lose some of the original industries in Table I when matching the NAICS 1997 codes to those of BEA 2002.
35
(1) (2) (3) (4)
Road Density x Input Narrowness 3.443*** 1.929*** 2.525*** 2.070***(0.652) (0.483) (0.434) (0.440)
Rule of Law x Contract Intensity 0.567*** 0.441***(0.056) (0.057)
Rule of Law x Input Narrowness 0.930 1.974**(0.866) (0.926)
log (Priv Credit/GDP) x Input Narrowness 2.344** 0.533(0.991) (1.082)
log GDP x Input Narrowness 0.464 0.897(0.721) (0.772)
Capital Intensity x log (K /L )c 0.007***(0.002)
Skill Intensity x log H c 0.008*(0.005)
Observations 13,179 14,231 26,157 20,823Rsquared 0.704 0.820 0.758 0.790Contract Intensity low high all allNumber of Countries 166 166 166 134Number of Industries 91 91 182 163Robust standard errors in parentheses. All regressions include country and industry fixed effects. The dependent variable is log(Expoc,k ) in year 2014.Contract intensity by industry measures are taken from Nunn (2007). The original measures are coded according to the NAICS 1997 classification andmatched to the BEA codes. Nunn (2007) measures contract intensity in k as the proportion of inputs of k classified as differentiated by Rauch (1999).Rule of law is from the World Governance Indicators (year 2014). Private credit over GDP is from the World Bank Indicators (averaged for 200514).GDP per capita, total GDP, the stock of physical capital (per capita), and the human capital index (H c ) are all taken from the Penn Tables (year 2014).Capital and skill intensity by industry are from the NBERCES Manuf. Industry Database (for year 2011). *** p<0.01, ** p<0.05, * p<0.1
TABLE VIRegressions on Industry Subsamples: Effects at Different Leves of Contract Intensity
how sensitive to e¢ cient contract enforcement industry k is, then the estimate in column
(1) should turn out to be signi�cantly milder than that one in column (2). The regressions
show, however, that the negative correlation is signi�cant in both subsamples, and moreover
the magnitude of the estimates are not signi�cantly di¤erent from one another.
Finally, to complement our previous results, columns (3) and (4) display the outcomes of two
regressions analogous, respectively, to those in columns (3) and (4) of Table II, but including
the interaction term between rule of law in country c and contract intensity of industry k.
Consistently with the previous results in the literature, the regressions show that countries
with better institutions exhibit a comparative advantage in the industries with high levels of
contract intensity. In addition to that, the regressions still support the prediction that countries
with denser road networks export relatively more in those sectors that need to source a larger
variety of intermediate inputs.
36
6 Further Discussion: Clustering and Coagglomeration
of Industries
The empirical results in the previous sections align with the model�s prediction that economies
with denser road networks exhibit a comparative advantage in industries with broad input bases.
The underlying mechanism of the model rests crucially on the idea that intermediate producers
(and industries in general) are unevenly distributed across space. As a consequence, those
industries that require a wide set of inputs will also be in higher need of costly transportation
for many of them. This section discusses the empirical plausibility of the notion that industries
using a large variety of inputs will be sourcing a vast number of them from di¤use locations, in
light of the recent evidence on clustering and coagglomeration of industries.
If the set of industries in the economy were uniformly distributed across space, then whether
a particular sector requires a wide or a narrow set of intermediate inputs would in principle have
no di¤erential e¤ect on their incurred transport costs. The di¤erential e¤ect of road density
across sectors with broad versus narrow input bases is, in fact, intrinsically linked to two features
of the geographic distribution of economic activities: i) concentration of speci�c industries in
certain geographic locations; ii) coagglomeration of industries with strong input-output links.
Both of these features have been vastly documented in di¤erent empirical studies.
Geographic Concentration by Industry: Examples of geographic concentration of speci�c
industries abound. Probably, the most often-cited examples of industry geographic concen-
tration are the high-tech �rms in Silicon Valley and the automobile industry in Detroit. In
addition to these paradigmatic cases, a growing number of articles have shown that industry
geographic clustering is quite a prevalent feature among a vast number of di¤erent industries,
and also in di¤erent countries. For example, Ellison and Glaeser (1997) study the degree of
concentration of U.S. manufacturing industries, and �nd that 446 out of 495 four-digit SIC
sectors display excess geographic concentration (relative to the degree of geographic concen-
tration that would be observed if �rms in all industries would pick their locations following an
identical random process). Moreover, they �nd that over one quarter of the U.S. manufacturing
industries exhibit what they consider �high geographic concentration�. Similar results are found
for France by Maurel and Sedillot (1999), for the U.K. by Devereux, Gri¢ th and Simpson
(2004) and by Duranton and Overman (2005), for Japan by Mori, Nishikimi and Smith (2005),
and for Belgium by Bertinelli and Decrop (2005). Our model in Sections 2 and 3 has assumed
a dispersed location of input sources. The rationale behind that assumption is to generate
geographic clustering of �rms by industry.
37
Coagglomeration of Strongly Linked Industries: One other crucial aspect of the model
leading to a di¤erential impact of road density on the cost of transportation in di¤erent in-
dustries is the (endogenous) location of �nal sectors next to their main input source. This
result of the model also aims at recreating a feature of the geographic distribution of industries
documented in the data. Ellison and Glaeser (1997) and Ellison, Glaeser and Kerr (2010) show
that the U.S. manufacturing industries with strong input-output links tend to locate in the
same geographic areas, whereas Duranton and Overman (2008) produce similar evidence for
the case of the U.K. Furthermore, Ellison, Glaeser and Kerr (2010) show that the presence of
input-output links between upstream and downstream industries represents the strongest factor
(amongst the Marshallian reasons for coagglomeration) leading to coagglomeration, and that
economizing on transport cost via physical proximity constitutes a key channel for this.
Input Narrowness and Geographic Concentration: An implicit feature in our model is
the heterogeneous surplus from coagglomeration across sectors di¤ering in their �j. In partic-
ular, the bene�t from coagglomeration between �nal good producers and intermediate inputs
increases as �j approaches either of the two extremes of the unit interval. In the model in the
main text 2.3 this speci�c feature does not arise in very transparent way, as the use of iceberg
transport cost that are linear in distance lead to corner solutions in (6), which depend only on
whether �j < 0:5 or �j > 0:5.40 Despite that, the knife-edge case when �j = 0:5 in (6) means
that �rms in that industry will be indi¤erent between location 0 and 1. As a consequence, one
can think that those �rms will choose amongst locations (possibly) in a random way. Bridging
this aspect to the data would imply that industries with smaller values of Ginij (i.e., those
with intermediate values of �j) will tend to be more dispersedly located than those with higher
values of Ginij (i.e., those with either low or high values of �j). Figure 3 shows that this fact
seems to be the veri�ed in the case of U.S. manufacturing industries.
Figure 3 displays a scatterplot showing the correlation between our measure of input nar-
rowness by industries and an index of their degree of geographic concentration. The index of
geographic concentration is taken from Ellison and Glaeser (1997).41 The higher the value of
the index, the more strongly concentrated the industry is. In addition, a value of the index
40See (33) and (34) in Appendix D for a case where the cut-o¤ value of �j depends also on other factors (e.g.,
wages, sectoral productivities, transport cost per unit of travelled distance). Recall also footnote 12 concerning
the e¤ect of convex cost in distance, and the possibility of interior solutions for sectors with values of �j that
lie far from the ends of the interval [0; 1]: Finally, the three-input case presented in Appendix D can also lead
to solutions where the intensity of agglomeration depends on the degree of disparity between the Cobb-Douglas
weights across the three intermediate goods, as showcased by (40).and (41).
41The index of geographic concentration in Ellison and Glaeser (1997) is computed for U.S. manufacturing
38
of geographic concentration above zero means that the industry exhibits stronger geographic
concentration than what would be observed if �rms within the industry chose their location
randomly. As the �gure shows, industries with narrower input bases tend to be geographically
more concentrated. This fact is consistent with the notion that industries with narrow input
bases are those that tend to bene�t more strongly from locating in close proximity to their
main intermediate input providers, and as a consequence show up in the data as geographically
more concentrated.
Figure 3: Geographic Concentration and Input Narrowness across Industries
industries classi�ed according to the 4-digit SIC codes. We matched the 4-digit SIC classi�cation to the BEA
industry classi�cation of the U.S. input-output matrix. Figure 3 plots the correlation between input narrowness
and geographic concentration using the BEA industry classi�cation.
39
7 Concluding Remarks
We proposed a simple trade model where the density of the internal transport network represents
a key factor in shaping comparative advantage and specialization. The underlying mechanism
rests on the idea that shipping intermediate inputs across spatial locations is costly. As a con-
sequence, industries that rely on a wide set of intermediate inputs become heavier users of the
internal transport network. The model shows that countries with denser transport infrastruc-
tures exhibit a comparative advantage in industries that combine a large variety intermediate
goods. Furthermore, when disparities in the internal transport network across countries are
su¢ ciently pronounced, they can sometimes even overturn more standard Ricardian patterns
of specialization driven by heterogeneities in labor productivities across sectors.
Drawing on intermediate goods transactions from the US input-output matrix to measure
industries�input breadth, we have also shown that the patterns of specialization predicted by
the model are broadly consistent with the trade �ows observed in the data. In particular, our
empirical analysis shows that countries with denser road networks tend to export relatively
more in industries that rely on a wide set of intermediate inputs.
Several caveats apply nonetheless to the empirical evidence. First, patterns of specialization
could be in�uencing investment in road infrastructure, and thus be behind the correlation found
in the data. In that respect, the fact that the same correlation appears when substituting
roadway density by waterway density seems reassuring, as waterways are harder to expand in
response to increased transport needs. Furthermore, our empirical results also remain in line
with the predictions of the model when we instrument the density of countries�road networks
with indicators of roughness of their terrain. Second, the measure of input breadth by industries
could alternatively be capturing a stronger need for contract enforcement when the input base
is wider. We showed however that the correlation predicted by our model is still present when
the confounding e¤ect of institutional quality (by country) and judiciary intensity (by industry)
is also taken into account. Finally, our measure of road density is a relatively imprecise way
to capture the e¢ ciency in connectedness of di¤erent locations. Road networks not only di¤er
in length, but also in terms of width, quality, etc. In addition, our measure of road density
fails to account for ine¢ ciencies in the layout of internal roads, and it also aggregates together
paved and unpaved roads. It would be certainly desirable to use of a more detailed and re�ned
measure of internal road networks by countries. Yet, it is hard envision a clear reason why the
above sources of measurement error in the road density indicator can end up systematically
biasing the previous empirical results in the direction predicted by the model.
40
Appendix A: Proofs and Additional Theoretical Results
Proof of Lemma 1. To prove that c�j(r1)=c�j(r2) � 1 for all �j 2 [0; 1] with strict inequality
if and only if �j 2 (0; 1), notice that '0(r) < 0 implies that 1 + '(r1)t > 1 + '(r2)t, and using(7) we have:
c�j(r1)=c�j(r2) =
([(1 + '(r1)t) = (1 + '(r2)t)]
�j for 0 � �j � 0:5;[(1 + '(r1)t) = (1 + '(r2)t)]
1��j for 0:5 � �j � 1:(24)
To prove the second part of the lemma, apply logs to c�j(r1)=c�j(r2) in (24), and di¤erentiate w.r.t.
�j, to obtain that @�ln�c�j(r1)=c
�j(r2)
��=@�j > 0 for 0 � �j < 0:5 and @
�ln�c�j(r1)=c
�j(r2)
��=@�j <
0 for 0:5 < �j � 1.
Proof of Proposition 1. We �rst prove by contradiction that ! = 1 cannot hold in
equilibrium. Given that the expression in (16) entails that total imports from F by H increase
with !, while total exports by H to F decrease with !, it will then follow that in equilibrium we
must necessarily have !� > 1, and that this equilibrium will be unique. We carry out the proof
of !� > 1 by splitting the possible parametric con�gurations of the model in three subsets.
i) Case 1: � � (1 + �)2. In this case, when ! = 1, using the LHS of (16), it follows that
total exports by H are equal to:
ExpoH =ln(1 + ")� ln(1 + �)2 ln(1 + ")� ln(�) : (25)
Notice that (1+�)2 � � implies the RHS of (25) is never greater than one half, while Assumption2 implies it is strictly above zero. Using now the RHS of (16), we can obtain that total imports
by H are:
ImpoH = 1�ln(1 + ") + ln(1 + �) + ln(�)
2 ln(1 + ") + ln(�): (26)
Comparing (25) versus (26), while bearing in mind � > 1, yields ExpoH > ImpoH . Hence,
when (1 + �)2 � �, the equilibrium must necessarily encompass ! > 1.
ii) Case 2: (1 + �)2 < � < (1 + ")2. Using setting again (16), we obtain:
ExpoH =ln(1 + ")� ln(1 + �) + ln(�)
2 ln(1 + ") + ln(�); (27)
while total imports by H are still given by (26). When (1 + �)2 < �, the RHS of (27) yields
a value strictly larger than one half, while the RHS of (26) is always strictly smaller than one
41
half. As a consequence, ExpoH > ImpoH also when (1+�)2 < � < (1+")2, and the equilibrium
must necessarily encompass ! > 1 in that range too.
iii) Case 3: (1 + ")2 < �. Using once again (16), notice that total exports by H are still
given by (27), which yields a value strictly above 0:5 and strictly below 1. In addition, total
imports by H are still given by (26), which yields a value strictly above 0, but strictly below
one half. Hence, when (1 + �)2 � �, the equilibrium must encompass ! > 1 as well.
Next, to prove that !� is strictly increasing in �, it su¢ ces to note that the boundaries in
all the expressions of the indicator functions If�g in (16) are all strictly increasing in �.
Lastly, to prove the di¤erent bounds on !� we proceed by contradiction for each of them.
First, suppose that (1 + ")2 > �. Notice that if !� � (1 + �)�0:5, then using (16) we can
observe that the mass of �nal goods exported by F would be at least one half. However, this is
incompatible with the fact that in equilibrium !� > 1. Hence, it must be that !� < (1+ �)�0:5.
Next, notice that when !� � (1 + ")=(1 + �), the exports by H fall to zero, while H�s imports
are strictly positive; hence, this cannot hold in equilibrium either, and it must be that !� <
(1 + ")=(1 + �). Second, suppose now that (1 + ")2 < �. In this case when !� � (1 + �)�1�0:5,exports by H would fall to zero, while H�s imports would still be strictly positive. Hence, it
must be the case that !� < (1 + �)�1�0:5.
Additional Results of Proposition 2. We plot below that additional two sub-cases
encompassed by Proposition 2 �case 1 (bis) and case 2 (bis).
Case 1 (bis) occurs when � < (1 + ")2 and the parametric con�guration of the model is such
that, in equilibrium, !� < (1 + �)�1 �12 . The graph is qualitatively similar to case 1 in Figure
1 of the main text, with the di¤erence that country H exports goods with �j 2 [0; �H), where�H lies above one half.
Case 2 (bis) takes place when � > (1 + ")2 and the parametric con�guration of the model
is such that, in equilibrium, !� > (1 + ")(1 + �). The main qualitative di¤erence between this
one and case 2 in Figure 1 is that here country F exports only those goods on the upper end
of [0; 1], namely those with �j 2 (�F ; 1].
42
Figure 1 (bis): Patterns of Trade and Specialization
43
A Simple Microfoundation of the Distance Function '(r)
We provide here a simple illustration of the geographical structure an the economy that serves
as microfoundation of the function '(r) assumed in Section 2.2.
Suppose there exists an infrastructure network connecting location 0 and 1 that comprise
two two types of pathways. One is a semi-circular path of total length �=2, which represents
the least direct path between the two locations. The other one is a road of length r 2 [0; 1],which allows shortening the distance between the two locations given by the semi-circular path.
The infrastructure network structure is plotted in Figure 4 for two di¤erent roads lengths,
namely 0 < r1 < r2 < 1. (Without any loss of generality, we will arbitrarily place the starting
point of the roads always in location 0.) The two extreme cases r = 0 and r = 1 would
correspond, respectively, to the case where the only path available from location 0 to location 1
is via the semi-circular arch of length �=2, and the case where the two locations are connected
by a straight horizontal line of length one.42
Figure 4: Geographic Structure of the Economy
42Conceptually, Figure 4 intends to represent the notion that road networks facilitate transportation across
location 0 and 1, relative to the (lengthier) semi-circular path of length �=2. Nothing precludes the fact that a
road network of length r could comprise several segments, whose lengths sum up to r. As it will become clear
next, this would not be optimal. More precisely, given a total length of road network equal to r 2 [0; 1], theshortest distance to connect inputs locations is achieved by building one single straight line.
44
Denoting by '(r) the shortest distance between location 0 and 1, given a road network of
length r 2 [0; 1], the following lemma characterizes the main properties of '(r).
Lemma 3 The shortest distance between location 0 and 1 is a strictly decreasing function of
the road length; i.e., '0(r) < 0. In addition, '00(r) < 0.43
Proof. To prove the lemma we �rst �nd the closed-form expression of the function '(r). Notice
that any straight segment connecting two points of the semi-circle linking location 0 to location
1 could be seen as a chord within a circle. We can then use the formula for the chord length in
circle, which in this case of only one straight segment of length r would state that
r = 2 � radius � sin�f � �2
�; (28)
where f equals the share of the whole semi-circle that is covered by the straight line of length
r.44 From (28), bearing in mind that radius = 0:5, it follows that the share of the semi-circle
covered by the road of length r is given by f = 2�� arcsin(r). Therefore, since the semi-circle
has total length equal to �=2, the total distance not covered by r is equal to �=2 � arcsin(r).Summing up to this last amount the total length of the road, r, we �nally obtain that:
'(r) =�
2+ r � arcsin(r); (29)
Next, di¤erentiating (29), we obtain
'0(r) = 1� (1� r2)� 12 ; (30)
which is strictly negative for any 0 < r � 1.Finally, di¤erentiating (30) , we can also observe that '00(r) = �r=(1� r2)� 3
2 , which is also
strictly negative for any 0 < r � 1. Finally, notice that '00(r) < 0 in turn implies that the
shortest way to link location 0 and 1 when the road length is r 2 [0; 1] is through one singlestraight segment of length r.
43The only crucial feature that the model needs is '0(r) < 0, which implies that road length lowers the cost
of internal transportation of goods. This is in fact the only assumption placed in the main text in Section 2.2.
The only implication of '00(r) < 0 for the model is that, for a given road length r, the shortest distance between
location 0 and 1 is achieved by building one single straight segment of length r, as plotted in Figure 3.44More formally, f equals the ratio between the angle formed by a straight line going from the center of the
semi-circle to the endpoint of the chord of length r, and an angle of 180�.
45
Appendix B: Additional Empirical Results
This appendix displays some additional empirical results that complement those in the main
text. For neatness, we split this appendix in two separate subsections.
Complementary Results for Section 4.3
This section �rst reports some summary statistics in Table A.1 corresponding to the main
variables of interest in Section 4.3. Next, Table A.2 and A.3 show the results of some regressions,
analogous to some of those previously presented in Section 4.3, but where the measure of the
density of the transport network is either changed or expanded. Finally, Table A.4 provides
some further robustness checks to the regressions in Section 4.3.
Variable Mean Std Dev Min Max Obs
Gini Coef. 0.93949 0.02314 0.88882 0.99342 294Herfindahl Index 0.10087 0.08036 0.02883 0.77590 294Coef. Var. 5.8063 2.03478 3.1955 17.2480 294LogVariance 8.4915 0.62126 9.5830 6.2111 294Roadways per sq km 0.78873 1.33132 0.00639 9.79747 166Railways per sq km 0.02148 0.02792 0.00007 0.13693 122Waterways per sq km 0.01360 0.02720 0.00005 0.17264 100
TABLE A.1Summary Statistics
(1) (2) (3) (4)
Railway Density x Input Narrowness 2.281*** 1.189*** 0.903*** 0.885***(0.133) (0.159) (0.162) (0.163)
Rule of Law x Input Narrowness 1.567** 2.408*** 2.359***(0.766) (0.824) (0.828)
log (Priv Credit/GDP) x Input Narrowness 4.328*** 2.048** 2.031**(0.936) (1.006) (1.007)
log GDP per capita x Input Narrowness 0.442 0.574 0.979(0.746) (0.782) (0.775)
Capital Intensity x log (K /L )c 0.009*(0.005)
Skil l Intensity x log H c 0.007***(0.001)
Observations 33,099 32,153 26,444 26,444Rsquared 0.754 0.756 0.794 0.793Number of Countries 122 118 109 109Number of Industries 294 294 259 259Robust standard errors in parentheses. All regressions include country and industry fixed effects. Railway density equals the total length of the railwaynetwork in km, divided by the area measured in sq km. Data of railway network length is taken from the CIA factbook. *** p<0.01, ** p<0.05, * p<0.1
TABLE A.2Transport Density measured by Railway Density
46
(1) (2) (3) (4) (5) (6)
Transp. (road + rail + waterway) Density x Input Nwness 7.268*** 2.680*** 1.519*** 4.108*** 2.376*** 1.881***(0.519) (0.598) (0.587) (0.302) (0.361) (0.375)
Rule of Law x Input Narrowness 2.381*** 2.892*** 2.394*** 3.091***(0.900) (0.936) (0.703) (0.764)
log (Priv Credit/GDP) x Input Narrowness 4.304*** 2.487** 3.537*** 1.698*(1.072) (1.114) (0.822) (0.913)
log GDP per capita x Input Narrowness 0.872 0.054 0.410 0.911(0.874) (0.933) (0.602) (0.656)
Capital Intensity x log (K /L )c 0.009 0.010**(0.006) (0.004)
Skil l Intensity x log H c 0.007*** 0.008***(0.002) (0.001)
Observations 25,180 24,234 21,014 42,578 40,692 31,892Rsquared 0.768 0.769 0.805 0.765 0.764 0.794Number of Countries 92 88 86 166 157 134Number of Industries 294 294 259 294 294 259Robust s tandard errors in parentheses . Al l regress ions include country and industry fixed effects . Columns (1) to (3) only include observationswhere informa tion on a l l transport measures (i .e., roadwa y, ra i lway and wa terway) i s ava i lable. Columns (4) to (6) a lso include observationswhere informa tion on ei ther ra i lway or wa terwa y (or both) a re miss ing, replacing the mi ss ing va lues by zero. *** p<0.01, ** p<0.05, * p<0.1
TABLE A.3Transport density measured by sum of roadways, railways and waterways per square km
In Table A.2, road density is replaced by railway density, as our main measure of depth of
local transport network. All the results follow a similar pattern as those in Section 4.3. In Table
A.3, we expand the measure of transport network density to include, in addition to roadways,
also railways and waterways. In this case, the density of the transport network of country c is
measured as the sum of total kilometers of roadways, railways and waterways, divided by the
area of the country. Again, all the results follow a similar pattern as those in Section 4.3.45
Table A.4 provides some �nal robustness checks, by restricting the samples of the regressions
reported in column (4) of Table 2 on a number of dimensions, and also by adding a few additional
covariates to that regression. Column (1) restricts the sample to countries with area greater
than 10,000 sq km, while column (2) restricts the sample to countries with population larger
than half million inhabitants. As we can see, results remain qualitatively unaltered when we
exclude small countries (either in size or population) from the sample. Column (3) excludes
very large countries in terms of their size. In particular, we drop from the sample countries
whose area is larger than 3,000,000 sq km.46 The rationale for this additional robustness check
45Note that for many countries we do not have information on railways or waterways, while we do have
information on roadways. This means that the total sample of the regressions in columns (1), (2) and (3) falls
substantially. For additional comparison, in columns (4), (5) and (6), we also include countries with missing
information on either railways or waterways (or in both), replacing the missing values by zeros.46This excludes the following seven countries from the sample: Russia, Canada, United States, China, Brazil,
Australia and India. The results are robust to setting the area-threshold for exclusion on alternative levels, such
47
(1) (2) (3) (4) (5) (6)
Road Density x Input Narrowness 2.589*** 1.767*** 2.034*** 1.584*** 1.733*** 1.748***(0.579) (0.399) (0.514) (0.485) (0.512) (0.511)
Rule of Law x Input Narrowness 2.869*** 3.929*** 3.172*** 3.363*** 3.350*** 2.888***(0.804) (0.833) (0.788) (0.769) (0.770) (0.842)
log (Priv Credit/GDP) x Input Narrowness 1.687* 0.793 1.611* 1.422* 1.600* 2.123**(0.950) (0.961) (0.928) (0.922) (0.940) (1.099)
log GDP per capita x Input Narrowness 0.889 0.797 0.983 0.736 5.417 0.207(0.676) (0.677) (0.663) (0.665) (5.263) (0.821)
Capital Intensity x log (K /L )c 0.009*** 0.011** 0.009** 0.010** 0.010** 0.017***(0.004) (0.004) (0.004) (0.004) (0.004) (0.005)
Skil l Intensity x log H c 0.007*** 0.006*** 0.007*** 0.008*** 0.008*** 0.012***(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
log area x Input Narrowness 0.606 0.572(0.397) (0.398)
log population x Input Narrowness 1.107** 3.599(0.453) (5.310)
log GDP x Input Narrowness 4.666(5.229)
Observations 29,760 30,085 30,817 31,892 31,892 24,467Rsquared 0.796 0.780 0.794 0.794 0.794 0.792Number of Countries 125 127 129 134 134 101Number of Industries 259 259 259 259 259 294Sample > 10 000 km2 > 500,000 pop < 3 mill ion km2 all countries all countries year roads 2010+Robust s tandard errors in parentheses . Al l regress ions include country and industry fi xed effects . Column (1) excludes countries whose area i s sma l ler than10,000 km2. Column (2) excludes countries whose population i s below 500,000 inhabitants . Column (3) excludes countries wi th area larger than 3,000,000 km 2.Column (6) excludes countries for which the measure of road dens i ty in the dataset wa s recorded before 2010. *** p<0.01, ** p<0.05, * p<0.1
TABLE A.4Additional Robustness Checks: area, population and year of roads in sample
is to account for the possibility that results may be a¤ected by the fact that some very large
countries may also have large swaths of uninhabitable land. Next, column (4) uses again the
entire sample of countries, but adds interactions terms between countries log area and Ginik,
and between log population and Ginik. This would control for the possibility that larger
countries may o¤er more opportunity for input diversity than smaller countries. Column (5)
includes an interaction term log GDP and Ginik, in case the aggregate size of the economy
may have some impact on specialization in sectors with di¤erent degrees of input narrowness.
Finally, column (6) excludes from the sample countries whose road network was measured
before year 2010, which corresponds to the median year in the sample (see details in Table A.8
in Appendix C). The results remain also qualitatively unaltered when restricting the sample to
countries whose road networks are more recently measured.
as at 7,000,000 km2 (which would leave India within the sample), or at 2,000,000 km2 (which would additionally
remove Argentina, Algeria, Congo, and Saudi Arabia from the sample).
48
Complementary Results for Section 5.1.1
Table A.5 shows the results of regressions analogous to those in Table IV in the main text, but
instead of splitting the sample of countries according to their income per head, it introduces a
triple interaction term between waterway density, the degree of input narrowness and the (log)
income per head of countries.
The results on Table A.5 are consistent with those of Table IV. In particular, we can see that
the triple interaction term carries always a positive and signi�cant coe¢ cient. This suggests
that the positive e¤ect of waterways density on specialization in industries with broad input
bases tends to be lower for richer economies than it is for lower-income countries.
(1) (2) (3) (4)
Waterways Density x Input Narrowness 1.634*** 1.921*** 2.037*** 2.155***(0.584) (0.606) (0.592) (0.614)
Waterways Density x Input Narrowness x log Income 0.486*** 0.554*** 0.738*** 0.702***(0.174) (0.180) (0.183) (0.189)
Rule of Law x Input Narrowness 4.968*** 4.731*** 3.420*** 3.811***(0.822) (0.853) (0.875) (0.904)
log (Priv Credit/GDP) x Input Narrowness 2.745** 1.037 2.793*** 1.076(1.071) (1.112) (1.071) (1.113)
log GDP per capita x Input Narrowness 1.174 0.235 1.052 0.145(0.854) (0.905) (0.854) (0.906)
Capital Intensity x log (K /L )c 0.007 0.007(0.006) (0.006)
Skil l Intensity x log H c 0.007*** 0.007***(0.001) (0.001)
Road Density x Input Narrowness 3.903*** 2.328***(0.757) (0.744)
Observations 25,975 22,357 25,975 22,357Rsquared 0.775 0.811 0.775 0.811Number of Countries 96 93 96 93Number of Industries 294 259 294 259Robust standard errors in parentheses. All regressions include country and industry fixed effects. The dependent variable is log(Expo c,k ) in year 2014.
Waterways data id taken from the CIA World Factbook, and comprises total length of navigable rivers, canals and other inland water bodies. Waterwaydensity equals internal waterways per square km. ***p<0.01, **p<0.05, *p<0.1
TABLE A.5Additional Robustness Checks: Waterways Density
49
Complementary Results for Section 5.1.2
Table A.6 shows in columns (1) - (3) the simple correlation between the di¤erent used measures
of terrain roughness in country c and road density in c. In all three cases the simple correlation
between the variables is negative and highly signi�cant. Next, in columns (4) - (6), we add
some additional country-level controls that may be a¤ecting road density (and which are used
in the regressions in the main text interacted with industry-level variables). As it can be readily
seen, the partial correlation between the three measures of terrain roughness and road density
always remains negative and highly signi�cant.
(1) (2) (3) (4) (5) (6)
Elevation Difference 0.020*** 0.012**(0.006) (0.005)
Std. Dev. Elevation 0.591*** 0.363**(0.184) (0.153)
% Mountainous 0.835*** 0.534**(0.296) (0.232)
Ruggedness
Rule of Law 0.371*** 0.329*** 0.404***(0.117) (0.112) (0.118)
Financial Development 0.045 0.038 0.046(0.133) (0.114) (0.098)
log Income 0.091 0.275 0.167(0.219) (0.204) (0.164)
Human Capital Index 0.028 0.076 0.099(0.202) (0.232) (0.237)
log (K/L)c 0.257 0.468*** 0.364***(0.188) (0.146) (0.134)
Observations 166 140 142 134 122 126Rsquared 0.100 0.051 0.037 0.306 0.366 0.383Robust standard reported errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1
TABLE A.6Terrain Roughness and Road Density
Dependent Variable: Roads per Km2
Table A.7 displays the results of a set of regressions that simultaneously include together as
independent variables rc � Ginik and terrain roughness interacted with Ginik. The rationalefor these regressions is to try to get some sense of whether, after controlling for the e¤ect
of the internal roadway network, terrain roughness may still display a systematic impact on
specialization across industries with di¤erent degrees of input narrowness. As it can be observed,
once the regressions control for the e¤ect of road density, the measures of terrain roughness tend
not to exhibit a signi�cant e¤ect on industry specialization.47 Table A.7 does not represent any
47The only exception is column (1), where the coe¢ cient is positive and signi�cant at 10%. This estimate
would imply that terrain roughness is associated with lower specialization in industries with wide input bases,
50
(1) (2) (3) (4) (5) (6)
Road Density x Input Narrowness 2.180*** 1.771*** 3.252*** 2.086*** 3.277*** 1.884***(0.380) (0.394) (0.521) (0.554) (0.516) (0.526)
Rule of Law x Input Narrowness 2.233*** 2.995*** 2.160*** 2.693*** 2.047*** 3.095***(0.707) (0.777) (0.729) (0.802) (0.739) (0.792)
Fin Dev x Input Narrowness 3.648*** 1.761** 4.174*** 2.728*** 4.051*** 1.600*(0.822) (0.910) (0.877) (0.953) (0.872) (0.937)
log GDP per capita x Input Narrowness 0.494 0.844 0.133 1.106 0.078 0.948(0.605) (0.665) (0.669) (0.720) (0.641) (0.675)
Capital Intensity x log (K /L )c 0.010** 0.009** 0.008*(0.004) (0.005) (0.004)
Skil l Intensity x log H c 0.008*** 0.007*** 0.008***(0.001) (0.001) (0.001)
Terrain Roughness x Input Narrowness 0.376* 0.186 1.331 1.601 0.015 0.027(0.210) (0.228) (1.293) (1.427) (0.022) (0.025)
Observations 40,692 31,892 35,988 29,229 36,544 30,067Rsquared 0.764 0.794 0.764 0.794 0.757 0.795Number of Countries 157 134 137 122 138 126Number of Industries 294 259 294 259 294 259Robust standard errors in parentheses. All regressions include country and industry fixed effects. 'Terrain Roughness' is measured in column (1) and (2) by thedifference between the max and min elevation within country c (source: CIA Factbook), in columns (3) and (4) by the std dev of elevation in country c at the 30''resolution (source: Ramcharan, 2009), and in columns (8) and (10) by the percentage of mountainous in terrain in country c (source: Fearon and Laitin, 2003).*** p<0.01, ** p<0.05, * p<0.1
TABLE A.7Direct Effect of Terrain Roughness
Measure of Terrain Roughness Elevation Difference Std. Dev. Elevation(CIA Wolrd Factbook) (Ramcharan, 2009)
% Mountainous Terrain(Fearon and Laitin, 2003)
sort of test about the validity of the exclusion restriction in the regressions in Table V. (In fact,
there is no way to test the validity of the exclusion restriction in the context of our paper.)
Yet, those results are comforting, in the sense that they somehow tame the concerns that the
instrument may be capturing some direct e¤ect of topography on specialization in industries
with broad input bases, besides the e¤ect mediated through its impact on road density.
even after controlling for the impact of road density. Notice, however, that the signi�cance disappears in column
(2), after we control for the e¤ect of factor endowments.
51
Appendix C: Further Data Details
country roads (km) year (roads) roads/sq km gdp pc (2014) country roads (km) year (roads) roads/sq km gdp pc (2014)Angola 51,429 2001 0.04125 7,968 Germany 645,000 2010 1.80661 45,961Albania 18,000 2002 0.62613 10,664 Djibouti 3,065 2000 0.13211 3,200UAE 4,080 2008 0.04880 64,398 Dominica 1,512 2010 2.01332 10,188Argentina 231,374 2004 0.08322 20,222 Denmark 74,497 2016 1.72871 44,924Armenia 7,792 2013 0.26198 8,586 Dominican Rep. 19,705 2002 0.40487 12,511Antigua & Barbuda 1,170 2011 2.64108 21,002 Algeria 113,655 2010 0.04772 12,812Australia 823,217 2011 0.10634 43,071 Ecuador 43,670 2007 0.15401 10,968Austria 133,597 2016 1.59289 47,744 Egypt 137,430 2010 0.13723 9,909Azerbaijan 52,942 2006 0.61134 15,887 Spain 683,175 2011 1.35183 33,864Burundi 12,322 2004 0.44276 772 Estonia 58,412 2011 1.29150 28,538Belgium 154,012 2010 5.04494 43,668 Ethiopia 110,414 2015 0.09999 1,323Benin 16,000 2006 0.14207 1,922 Finland 454,000 2012 1.34262 40,401Burkina Faso 15,272 2010 0.05570 1,565 Fiji 3,440 2011 0.18825 7,909Bangladesh 21,269 2010 0.14326 2,885 France 1,028,446 2010 1.59746 39,374Bulgaria 19,512 2011 0.17598 17,462 Gabon 9,170 2007 0.03426 14,161Bahrain 4,122 2010 5.42368 41,626 United Kingdom 394,428 2009 1.61910 40,242Bahamas, The 2,700 2011 0.19452 23,452 Georgia 19,109 2010 0.27416 9,362Bosnia and Herz. 22,926 2010 0.44780 10,028 Ghana 109,515 2009 0.45912 3,570Belarus 86,392 2010 0.41615 20,290 Guinea 44,348 2003 0.18038 1,429Belize 2,870 2011 0.12497 8,393 Gambia, The 3,740 2011 0.33097 1,544Bermuda 447 2010 8.27778 57,531 GuineaBissau 3,455 2002 0.09564 1,251Bolivia 80,488 2010 0.07327 6,013 Equatorial Guinea 2,880 2000 0.10267 40,133Brazil 1,580,964 2010 0.18565 14,871 Greece 116,960 2010 0.88635 25,990Barbados 1,600 2011 3.72093 14,220 Grenada 1,127 2001 3.27616 11,155Bhutan 10,578 2013 0.27551 6,880 Guatemala 17,332 2015 0.15917 6,851Central African Rep. 20,278 2010 0.03255 594 Hong Kong 2,100 2015 1.89531 51,808Canada 1,042,300 2011 0.10439 42,352 Honduras 14,742 2012 0.13152 4,424Switzerland 71,464 2011 1.73133 58,469 Croatia 26,958 2015 0.47634 21,675Chile 77,764 2010 0.10285 21,581 Haiti 4,266 2009 0.15373 1,562China 4,106,387 2011 0.42788 12,473 Hungary 203,601 2014 2.18860 25,758Cote d'Ivoire 81,996 2007 0.25428 3,352 Indonesia 496,607 2011 0.26075 9,707Cameroon 51,350 2011 0.10801 2,682 India 4,699,024 2015 1.42946 5,224Congo, Dem. Rep. 153,497 2004 0.06546 1,217 Ireland 96,036 2014 1.36661 48,767Congo, Rep. 17,000 2006 0.04971 4,426 Iran 198,866 2010 0.12066 15,547Colombia 204,855 2015 0.17987 12,599 Iraq 59,623 2012 0.13603 12,096Comoros 880 2002 0.39374 1,460 Iceland 12,890 2012 0.12515 42,876Cabo Verde 1,350 2013 0.33474 6,290 Israel 18,566 2011 0.89389 33,270Costa Rica 39,018 2010 0.76356 14,186 Italy 487,700 2007 1.61844 35,807Curacao 550 N.A. 1.23874 25,965 Jamaica 22,121 2011 2.01265 7,449Cayman Islands 785 2007 2.97348 51,465 Jordan 7,203 2011 0.08062 10,456Cyprus 20,006 2011 2.16258 28,602 Japan 1,218,772 2015 3.22499 35,358Czech Republic 130,661 2011 1.65673 31,856 Kazakhstan 97,418 2012 0.03575 23,450
TABLE A.8
52
country roads (km) year (roads) roads/sq km gdp pc (2014) country roads (km) year (roads) roads/sq km gdp pc (2014)Kenya 160,878 2013 0.27720 2,769 Paraguay 32,059 2010 0.07882 8,284Kyrgyzstan 34,000 2007 0.17004 3,359 Qatar 9,830 2010 0.84844 144,340Cambodia 44,709 2010 0.24696 2,995 Romania 84,185 2012 0.35314 20,817Korea, South 104,983 2009 1.05278 35,104 Russia 1,283,387 2012 0.07506 24,039Kuwait 6,608 2010 0.37086 63,886 Rwanda 4,700 2012 0.17845 1,565Laos 39,586 2009 0.16717 5,544 Saudi Arabia 221,372 2006 0.10298 48,025Lebanon 6,970 2005 0.67019 13,999 Sudan 11,900 2000 0.00639 3,781Liberia 10,600 2000 0.09518 838 Senegal 15,000 2015 0.07625 2,247Sri Lanka 114,093 2010 1.73896 10,342 Singapore 3,425 2012 4.91392 72,583Lithuania 84,166 2012 1.28891 28,208 Sierra Leone 11,300 2002 0.15751 1,419Latvia 72,440 2013 1.12155 23,679 El Salvador 6,918 2010 0.32879 7,843Morocco 58,395 2010 0.13077 7,163 Serbia 44,248 2010 0.57113 13,441Moldova 9,352 2012 0.27627 4,811 Sao Tome & Princ. 320 2000 0.33195 3,239Madagascar 37,476 2010 0.06384 1,237 Suriname 4,304 2003 0.02627 15,655Maldives 88 2013 0.29530 14,391 Slovakia 54,869 2012 1.11898 28,609Mexico 377,660 2012 0.19225 15,853 Slovenia 38,985 2012 1.92300 30,488Macedonia 14,182 2014 0.55155 13,151 Sweden 579,564 2010 1.28708 44,598Mali 22,474 2009 0.01812 1,434 Seychelles 526 2015 1.15604 25,822Malta 3,096 2008 9.79747 31,644 Syria 69,873 2010 0.37732 4,200Burma 34,377 2010 0.05081 5,344 Turks and Caicos 121 2003 0.12764 20,853Montenegro 7,762 2010 0.56198 14,567 Chad 40,000 2011 0.03115 2,013Mongolia 49,249 2013 0.03149 11,526 Togo 11,652 2007 0.20520 1,384Mozambique 30,331 2009 0.03794 1,137 Thailand 180,053 2006 0.35090 13,967Mauritania 10,628 2010 0.01031 3,409 Tajikistan 27,767 2000 0.19269 2,747Mauritius 2,149 2012 1.05343 17,942 Turkmenistan 58,592 2002 0.12004 20,953Malawi 15,450 2011 0.13040 949 Trinidad & Tobago 9,592 2015 1.87051 31,196Malaysia 144,403 2010 0.43779 23,158 Tunisia 19,418 2010 0.11868 10,365Niger 18,949 2010 0.01496 852 Turkey 385,754 2012 0.49231 19,236Nigeria 193,200 2004 0.20914 5,501 Tanzania 86,472 2010 0.09128 2,213Nicaragua 23,897 2014 0.18330 4,453 Uganda 20,000 2011 0.08297 1,839Netherlands 138,641 2014 3.33729 47,240 Ukraine 169,694 2012 0.28116 10,335Norway 93,870 2013 0.28990 64,274 Uruguay 77,732 2010 0.44112 20,396Nepal 10,844 2010 0.07368 2,173 United States 6,586,610 2012 0.66981 52,292New Zealand 94,902 2012 0.35301 34,735 Uzbekistan 86,496 2000 0.19333 8,195Oman 60,230 2012 0.19460 38,527 Venezuela 96,189 2014 0.10546 14,134Pakistan 263,942 2014 0.33155 4,646 British Virgin Isl . 200 2007 1.32450 26,976Panama 15,137 2010 0.20070 19,702 Vietnam 195,468 2013 0.59016 5,353Peru 140,672 2012 0.10945 10,993 Yemen 71,300 2005 0.13505 3,355Philippines 216,387 2014 0.72129 6,659 South Africa 747,014 2014 0.61276 12,128Poland 412,035 2012 1.31773 25,156 Zambia 40,454 2005 0.05375 3,726Portugal 82,900 2008 0.90021 28,476 Zimbabwe 97,267 2002 0.24892 1,869
TABLE A.8 (cont.)
53
Appendix D: Robustness of Theoretical Results
1. Additive Linear Transport Cost of Intermediate Inputs
The benchmark model has used the assumption of multiplicative iceberg transport cost that
increase linearly with distance. Suppose instead that transport cost of input i to location j is
linearly additive in distance. That would imply that, given lj 2 [0; 1], producer j faces
p0;j = (1 + "0)w + lj'(r)t and p1;j = (1 + "1)w + (1� lj)'(r)t
As a result, producer j will choose his optimal location by solving:
minlj2[0;1]
: ln [cj(lj)] = (1� �j) ln [(1 + "0)w + lj'(r)t] + �j ln [(1 + "1)w + (1� lj)'(r)t] ; (31)
where (31) is using the logarithm of cj(lj) for algebraic convenience. Computing the second
derivative of ln [cj(lj)] with respect to lj yields:
@2 ln [cj(lj)]
(@lj)2= � (1� �j) ('(r)t)2
[(1 + "0)w + lj'(r)t]2 �
�j ('(r)t)2
[(1 + "1)w + (1� lj)'(r)t]2< 0: (32)
Therefore the FOC of (31) would yield a maximum instead of a minimum. Also, due to (32),
the solution of (31) would always be a corner solution, with either l�j = 0 or l�j = 1:
Despite being given by a corner solution, the optimal location choices by �nal producers
di¤er slightly from those in (6), because in the case of additive transport cost the values of
"0 and "1 also a¤ect the optimal location of j. To see this, let e� be de�ned by the value of�j 2 (0; 1) that solves the following equation
[(1 + "0)w]1�e�
[(1 + "1)w]e� =
[(1 + "0)w + '(r)t]1�e�
[(1 + "1)w + '(r)t]e� : (33)
The optimal location by producer j is given by
l�j =
8<: 0 if �j � e�1 if �j � e� : (34)
Using the result in (34), we can also obtain that
c�j =
8<: [(1 + "0)w]�j [(1 + "1)w + '(r)t]
1��j if �j � e�[(1 + "0)w + '(r)t]
�j [(1 + "1)w]1��j if �j � e� ;
from where an analogous result to that one Lemma 1 follows.
Finally, as a last remark, notice that in the special case in which "0 = "1, we have that (33)
yields e� = 0:5, and thus the optimal location choice (34) becomes identical to that in (6).54
2. General CES Production Function
Here we show that the main results of Section 2 hold true when (3) is replaced by a more general
CES production function. We let now total output of �nal good j 2 [0; 1] be represented by:
Yj = [(1� �j) X�0;j + �j X
�1;j]
1=�; with �j 2 [0; 1] and � < 1:
In this case, it can be shown that the marginal cost of producing good j is given by:
cj =
"(1� �j)1=(1��)
p�=(1��)0;j
+�1=(1��)j
p�=(1��)1;j
#�(1��)=�: (35)
Producer j will choose lj 2 [0; 1] in order to minimize cj, bearing in mind input prices asfunctions of lj; namely p0;j = [1 + lj'(r)t] (1 + "0)w and p1;j = [1 + (1� lj)'(r)t] (1 + "1)w.De�ning:
�(lj) �(1� �j)1=(1��)
[1 + lj'(r)t]�=(1��) (1 + "0)
�=(1��) +�1=(1��)j
[1 + (1� lj)'(r)t]�=(1��) (1 + "1)�=(1��);
notice that �nding the lj 2 [0; 1] that minimizes (35) is an isomorphic to maximizing �(lj).Furthermore, computing the second derivative of �(lj) with respect to lj yields:
@2�
(@lj)2 =
� ('(r)t)2
(1� �)2
8<:(1� �j)1
1��
(1 + "0)�
1��[1 + lj'(r)t]
� 11���1 +
�1
1��j
(1 + "1)�
1��[(1� lj)'(r)t]�
11���1
9=; > 0:
Therefore, the solution of �0 (lj) = 0 would yield a minimum for �(lj), which implies that it
would yield a maximum for cj. This in turn implies that the minimum of cj must be on a
corner solution; that is, either l�j = 0 or l�j = 1:
Computing now cj(0) and cj(1) from (35), together with the expressions for p0;j and p1;j,
and comparing them we can then obtain:
l�j =
8><>:0 if �j �
h1 +
�1+"11+"0
��i�11 if �j �
h1 +
�1+"11+"0
��i�1 ; (36)
where notice that (36) boils down to (6) in the case of � = 0 (i.e., the Cobb-Douglas production
case). Finally, combining (36) together with (35), we can observe that a result analogous to
Lemma 1 obtains in this generalized case as well.
55
3. An Example with Three Intermediate Inputs
We provide now an example with three intermediate goods located in three di¤erent locations.
(This example could be generalized to a case with N intermediate goods.) We denote the
intermediate goods by the letters A;B and C. Each input is produced in a speci�c location,
which label with the same letter as that of the input. To keep the analysis simple and brief,
we assume that all three locations are equidistant from each other.
There exist three subsets of �nal goods, g = A;B;C. Each subset g comprises a continuum
of �nal goods j. The production function of �nal good j in each subset are given by:
g = A: YA;j =h�
13+ j
� 13+ j�13� j
2
� 23� ji�1
X13+ j
A;j X13� j
2B;j X
13� j
2C;j , (37)
g = B: YB;j =h�
13+ j
� 13+ j�13� j
2
� 23� ji�1
X13� j
2A;j X
13+ j
B;j X13� j
2C;j , (38)
g = C: YC;j =h�
13+ j
� 13+ j�13� j
2
� 23� ji�1
X13� j
2A;j X
13� j
2B;j X
13+ j
C;j ; (39)
where j 2�0; 2
3
�:
The expressions in (37)-(39) imply that �nal goods in subset g = A tend to use input A
relatively more intensively, and so on and so forth forth for g = B and g = C. Also, note
that the degree of intensity of �nal good j in its most important intermediate input grows
with j, and Yg;j = Xg;j, for g = A;B;C, when j =23. Finally, notice that when j = 0,
all intermediate inputs are equally important. In that regard, the case with j =23would be
analogous to the cases with either �j = 0 or �j = 1 in the main text, whereas the case with
j = 0 would be analogous to the case where �j =12.
Consider a generic good j of the subset g = A. (The results for g = B and g = C can be
obtained by appropriately relabeling the variables.) The marginal cost of production would be
cA;j = p13+ j
A;j p13� j
2B;j p
13� j
2C;j :
Let lj;A 2 [0; 1] denote now the distance from the location chosen by this generic �rm j to
the location of intermediate input A. A complete analysis of the optimal location choice would
require a thourough description of the geographic setup of the economy. In what follows, to
keep the analysis brief, we show two simple cases under two alternative assumptions concern-
ing location bilateral distances. We dub these cases as: i) straight-line bilateral choices; ii)
diagonal-line choices.
1. Straight-line bilateral location choices: Assume that when lj;A = 0, then lj;B = lj;C = 1.
On the other hand, when 0 < lj;A � 1, we assume that lj;B = 1 � lj;A and lj;C = 1.48
48Nothing would change if we instead assumed that, when 0 < lj;A � 1; lj;C = 1� lj;A and lj;B = 1.
56
Hence, a �rm producing �nal good j of the subset g = A would pick its location by solving
minlj;A2[0;1]
: ln [cA;j (lj;A)] =�13+ j
�ln (1 + lj;A'(r)t)+
�13� 1
2 j�ln [1 + (1� lj;A)'(r)t]+�;
where � is a constant. This problem yields again corner solutions, where l�j;A = 0 whenever
j 2�0; 2
3
�, while when j = 0 the �rm is indi¤erent between any of the locations where
intermediate goods are produced. In addition, a result analogous to that in Lemma 1
obtains: the cost-reduction e¤ect of r is proportionally larger for sectors with small j.
2. Diagonal-line location choices: Assume that when lj;A = 0, then lj;B = lj;C = 1. On the
other hand, when 0 < lj;A � 1, we assume that lj;B = 1� 12lj;A and lj;C = 1� 1
2lj;A.49 In
this case, a �rm producing �nal good j of the subset g = A would pick its location by
solving
minlj;A2[0;1]
: ln [cA;j (lj;A)] =�13+ j
�ln (1 + lj;A'(r)t)+2
�13� 1
2 j�ln�1 +
�1� 1
2lj;A�'(r)t
�+�;
where � is a constant. This problem yields again corner solutions, where in the optimum
either l�j;A = 0 or l�j;A = 1. However, di¤erent from straight-line bilateral case above, in
this diagonal-line case we have that there may exist values of j > 0 for which in the
optimum l�j;A = 1. In particular,
l�j;A =
(0; if j � maxf0;g1; if j � maxf0;g
(40)
where,
� 1
3
2 ln�[1 + '(r)t] =
�1 + 1
2'(r)t
�� ln (1 + '(r)t)
ln�[1 + '(r)t] =
�1 + 1
2'(r)t
�+ ln (1 + '(r)t)
: (41)
Notice from the expression in (41) that there may be feasible parametric con�gurations
for which > 0. On the other hand, it must necessarily be the case that < 13.
One important di¤erence implied by (40) is that the exact optimal location of �nal good
producer j not only depends now on the value of j, but also (possibly) on '(r) via .
In any case, while the results are more nuanced than in the straight-line case, they keep
the same qualitative features. Furthermore, an analogous result to Lemma 1 still applies
to this case as well.
49This implicitly restricts the location choice of a �nal good producer of subset A to never set either lj;C < 0:5
or lj;B < 0:5. Nevertheless, given the structure of (37), this would never be optimal in this case.
57
References
[1] Acemoglu, D., Antràs, P. and Helpman, E. (2007). �Contracts and Technology Adoption�, Amer-
ican Economic Review, vol. 97, pp. 916-943.
[2] Acemoglu, D. and Zilibotti, F. (1997). �Was Prometheus Unbound by Chance? Risk, Diversi�-
cation and Growth�, Journal of Political Economy, vol. 105, pp. 709-751.
[3] Agnosteva, D., Anderson, J. and Yotov, Y. (2014). �Intra-national Trade Costs: Assaying Regional
Frictions�, NBER WP No. 19872.
[4] Allen, T. and Arkolakis, C. (2014). �Trade and the Topography of the Spatial Economy�, Quarterly
Journal of Economics, vol. 129, pp. 1085�1140.
[5] Anderson, J. E., and Van Wincoop, E. (2004). �Trade costs�, Journal of Economic Literature,
vol. 42, pp. 691�751.
[6] Antràs, P. (2005). �Incomplete Contracts and the Product Cycle�, American Economic Review,
vol. 95, pp. 1054-1073.
[7] Atkin, D. and Donaldson, D. (2015). �Who�s Getting Globalized? The Size and Implications of
Intra-national Trade Costs�, Mimeo, MIT.
[8] Barro, R. and Lee, J. (2013). �A New Data Set of Educational Attainment in the World, 1950-
2010�, Journal of Development Economics, vol 104, pp.184-198.
[9] Beck, T. (2002). �Financial Development and InternationalTrade. Is There a Link?�, Journal of
International Economics, vol. 57, pp. 107-131.
[10] Becker, B., Chen, J. and Greenberg, D. (2012). �Financial development, �xed costs, and interna-
tional trade�, Review of Corporate Finance Studies, vol. 2, pp. 1-28.
[11] Becker, R., Gray, W. and Marvakov, J. (2013). �NBER-CES manufacturing industry database:
Technical notes�, NBER.
[12] Bertinelli, L. and Decrop, J. (2005). �Geographical agglomeration: Ellison and Glaeser�s index
applied to the case of Belgian manufacturing industry�, Regional Studies, vol. 39, pp. 567-583
[13] Blanchard, O. and Kremer, M. (1997). �Disorganization�, Quarterly Journal of Economics, Vol.
112, pp. 1091-1126.
58
[14] Chor, D. (2010). �Unpacking sources of comparative advantage: A quantitative approach�, Journal
of International Economics, vol. 82, pp. 152-167.
[15] Clague, C. (1991a). �Relative E¢ ciency, Self-Containment, and Comparative Costs of Less De-
veloped Countries�, Economic Development and Cultural Change, Vol. 39, pp. 507-530.
[16] Clague, C. (1991b). �Factor Proportions, Relative E¢ ciency and Developing Countries�Trade�,
Journal of Development Economics, Vol. 35, pp. 357-380.
[17] Cosar, K. and Fajgelbaum, P. (2016). �Internal Geography, International Trade, and Regional
Specialization�, American Economic Journal: Microeconomics, vol. 8, pp. 24-56.
[18] Costinot, A. (2009), �On the Origins of Comparative Advantage�, Journal of International Eco-
nomics, vol. 77, pp 255-264.
[19] Cuñat, A. and Melitz, M. (2012). �Volatility, labor market �exibility, and the pattern of compar-
ative advantage�, Journal of the European Economic Association, vol. 10, pp 225-254.
[20] Devereux, M., Gri¢ th, R. and Simpson, H. (2004). �The geographic distribution of production
activity in the UK�, Regional Science and Urban Economics, vol. 34, pp. 533-564.
[21] Donaldson, D. (2018). �Railroads of the Raj: Estimating the Economic Impact of Transportation
Infrastructure�, American Economic Review, vol. 108, pp. 899-934.
[22] Donaldson, D. and Hornbeck, R. (2016). �Railroads and American Economic Growth: A "Market
Access" Approach�, Quarterly Journal of Economics, vol. 131, pp 799-858.
[23] Dornbusch, R., Fischer, S. and Samuelson, P. (1977). �Comparative advantage, trade, and pay-
ments in a ricardian model with a continuum of goods�, American Economic Review, vol. 67.
[24] Duranton, G., Morrow, P. and Turner, M. (2014). �Roads and Trade: Evidence from the US�,
Review of Economic Studies, vol. 81, pp. 681-724.
[25] Duranton, G. and Overman, H. (2005). �Testing for Localization Using Micro-Geographic Data�,
Review of Economic Studies, Vol. 72, pp. 1077�1106.
[26] Duranton, G. and Overman, H. (2008). �Exploring the detailed location patterns of UK manufac-
turing industries using microgeographic data�, Journal of Regional Science, vol. 48, pp. 213�243.
[27] Ellison, G., Glaeser, E. and Kerr. W. (2010). �What Causes Industry Agglomeration? Evidence
from Coagglomeration Patterns�, American Economic Review, vol. 100, pp. 1195-1213.
59
[28] Ellison, G. and Glaeser, E. (1997). �Geographic concentration in US manufacturing industries: a
dartboard approach�, Journal of Political Economy, vol. 105, pp. 889-927.
[29] Fajgelbaum, P. and Redding, S. (2014). �External Integration, Structural Transformation and
Economic Development: Evidence from Argentina 1870-1914�, NBER WP No. 20217.
[30] Fearon, J. and Laitin, D. (2003). �Ethnicity, Insurgency, and Civil War�, American Political
Science Review, vol. 97, pp. 75�90.
[31] Ferguson, S. and Formai, S. (2013) �Institution-Driven Comparative Advantage and Organiza-
tional Choice�, Journal of International Economics, vol. 90, pp. 193-200.
[32] Gaulier, G. and Zignago, S. (2010). �BACI: International Trade Database at the Product-level.
The 1994-2007 Version�, CEPII Working Paper 2010-23.
[33] Greenwood, J. and Jovanovic, B. (1990). �Financial Development, Growth, and the Distribution
of Income�, Journal of Political Economy, vol. 98, pp. 1076-1107.
[34] Hillberry, R. and Hummels, D. (2008). �Trade responses to geographic frictions: A decomposition
using micro-data�, European Economic Review, vol. 52, pp. 527-550.
[35] Imbs, J. and Wacziarg, R. (2003). �Stages of Diversi�cation�, American Economic Review, vol.
93, pp. 63-86.
[36] Levchenko, A. (2007). �Institutional Quality and International Trade�, Review of Economic Stud-
ies, vol. 74, pp. 681-724.
[37] Limao, N. and Venables, A. (2001). �Infrastructure, geographical disadvantage, transport costs,
and trade�, World Bank Economic Review, vol. 15, pp. 451�479.
[38] Manova, K. (2013). �Credit Constraints, Heterogeneous Firms, and International Trade�, Review
of Economic Studies, vol. 80, pp. 711-744.
[39] Maurel, F. and Sedillot, B. (1999). �A measure of the geographic concentration in french manu-
facturing industries�, Regional Science and Urban Economics, vol. 29, pp. 575-604.
[40] Matsuyama, K. (2017). �Geographical Advantage: Home Market E¤ect in a Multi-Region World�,
Research in Economics, vol. 71, pp. 740-758.
[41] Mesquita Moreira M., Blyde, J., Volpe, C. and Molina, D. (2013). �Too Far to Export: Domestic
transport costs and regional export disparities in latin america and the caribbean�, IADB.
60
[42] Mori, T., Nishikimi, K. and Smith, T. (2005). �A Divergence Statistic for Industrial Localization�,
Review of Economics and Statistics, vol. 87, pp. 635-651.
[43] Nunn, N. (2007). �Relationship-speci�city, incomplete contracts, and the pattern of trade�, Quar-
terly Journal of Economics, vol. 122, pp.569-600.
[44] Nunn, N. and Puga, D. (2012). �Ruggedness: The blessing of bad geography in Africa�, Review
of Economics and Statistics, vol. 94, pp.20-36.
[45] Ramcharan, R. (2009). �Why an economic core: domestic transport costs�, Journal of Economic
Geography, vol. 9, pp. 559-581.
[46] Ramondo, N., Rodriguez-Clare, A. and Saborio-Rodriguez, M. (2016). �Trade, Domestic Frictions,
and Scale E¤ects�, American Economic Review, vol. 106, pp. 3159-84.
[47] Rauch, J. (1999). �Networks versus markets in international trade�, Journal of International
Economics, vol. 48, pp. 7-35.
[48] Redding, S. (2016). �Goods Trade, Factor Mobility and Welfare�, Journal of International Eco-
nomics, vol. 101, pp. 148-167.
[49] Riley, S., DeGloria, S. and Elliot, R. (1999) �A Terrain Ruggedness Index That Quanti�es Topo-
graphic Heterogeneity�, Intermountain Journal of Sciences, vol. (5), pp. 23�27.
[50] Saint-Paul, G. (1992), �Technological Choice, Financial Markets and Economic Development�,
European Economic Review, vol. 36, pp. 763-781.
[51] Svaleryd, H. and Vlachos, J. (2005) �Financial markets, the pattern of industrial specialization
and comparative advantage: Evidence from OECD countries�, European Economic Review, vol.
49, pp. 113-144.
[52] Volpe Martincus, C. and Blyde, J. (2013). �Shaky roads and trembling exports: Assessing the
trade e¤ects of domestic infrastructure using a natural experiment�, Journal of International
Economics, vol. 90, pp. 148-161.
[53] Volpe Martincus, C., Carballo, J. and Cusolito, A. (2017). �Roads, exports and employment:
Evidence from a developing country�, Journal of Development Economics, vol. 125, pp. 21-39.
[54] Yeaple, S. and Golub, S. (2007). �International productivity di¤erences, infrastructure, and com-
parative advantage�, Review of International Economics, vol. 15, pp. 223-242.
61